CN116304337A - Object recommendation model training method, recommendation object determining method and device - Google Patents

Abstract

The embodiment of the specification provides an object recommendation model training method, a recommendation object determining method and a recommendation object determining device. And determining the user characteristic embedded representation of the user characteristic and the object characteristic embedded representation of the target object and the user interaction object during each round of model training. Decoupling the user interest embedded characterization and the user popular embedded characterization from the object feature embedded characterization of the user interaction object; determining target object characterization, user interest characterization and user popular characterization according to the object feature embedded characterization of the target object and the coupling result between the user feature embedded characterization and the user interest embedded characterization and the user popular embedded characterization respectively; and adjusting model parameters of the object recommendation model according to the user interest prediction results determined based on the user interest characterization, the user secondary characterization and the target object characterization and the user secondary prediction results.

Description

Object recommendation model training method, recommended object determination method and device

Technical Field

The embodiments of this specification generally relate to the field of object recommendation, and more particularly to an object recommendation model training method, a recommended object determination method and device.

Background Art

With the development of artificial intelligence technology, object recommendation schemes based on object recommendation models are increasingly being applied to object recommendation scenarios. Object recommendation models are a type of machine learning model. Object recommendation models can usually be modeled based on sequential behavior or based on decoupled learning. Common modeling schemes based on sequential behavior usually only learn embedding representations dominated by user interests based on historical behavior, without distinguishing between potential behavioral intentions. Common modeling schemes based on decoupled learning are difficult to explicitly include user intentions such as user conformity and user risk preferences when recommending objects. In addition, due to the lack of labeled data to explicitly constrain these user intentions, it is difficult to well characterize these user intentions in modeling. The above-mentioned defects of the modeling schemes based on sequential behavior and the modeling schemes based on decoupled learning make the above-mentioned modeling schemes based on sequential behavior and the modeling schemes based on decoupled learning unsuitable for application in some object recommendation scenarios, such as commodity intelligent recommendation scenarios, fund intelligent recommendation scenarios, etc.

Summary of the invention

The embodiments of this specification provide an object recommendation model training method, a recommended object determination method and a device. Using the object recommendation model training scheme, an object recommendation model can be built based on user interests, user herd mentality and user risk preference, thereby improving the user understanding effect of the object recommendation model.

According to one aspect of an embodiment of the present specification, a method for training an object recommendation model is provided, wherein the object recommendation model is used to determine whether to recommend a target object to a user, the method comprising: looping through the following model training process until a training end condition is met: providing user features, object features of user interaction objects, and object features of target objects to an embedding layer of the object recommendation model, respectively, to obtain a user feature embedding representation of the user features, an object feature embedding representation of the user interaction objects, and an object feature embedding representation of the target object; decoupling, via a decoupling layer of the object recommendation model, a user interest embedding representation and a user interest embedding representation corresponding to user interests and user herd mentality, respectively, from the object feature embedding representation of the user interaction objects The invention relates to a method for preparing a user conformity embedding representation; fusing the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, and providing the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation and the target object representation; predicting the user interest prediction result based on the user interest representation and the target object representation, and predicting the user conformity prediction result based on the user conformity representation and the target object representation; and in response to the training end condition not being met, determining the total loss function based on the user interest prediction result and the user conformity prediction result, and adjusting the model parameters of the object recommendation model according to the total loss function.

Optionally, in an example of the above aspect, the method may also include: providing the object feature embedding representations of the target object and the user interaction object to the graph neural network of the object recommendation model respectively to perform graph enhancement processing based on the object knowledge graph, so as to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing.

Optionally, in an example of the above aspect, the user interaction object includes at least two user interaction objects. Decoupling the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of the user interaction object via the decoupling layer of the object recommendation model may include: decoupling the object feature embedding representation components corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of each user interaction object via the decoupling layer of the object recommendation model and cascading them to obtain the user interest embedding representation and the user conformity embedding representation.

Optionally, in an example of the above aspect, decoupling a user interest embedding representation and a user conformity embedding representation corresponding to user interest and user conformity respectively from the object feature embedding representation of the user interaction object via the decoupling layer of the object recommendation model may include: decoupling a user interest embedding representation and a user conformity embedding representation corresponding to user interest and user conformity respectively from the object feature embedding representation of the user interaction object based on an unsupervised mechanism via the decoupling layer of the object recommendation model.

Optionally, in an example of the above aspect, the decoupling layer includes a self-attention network. Decoupling the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of the user interaction object via the decoupling layer of the object recommendation model may include: providing the object feature embedding representation of the user interaction object to the self-attention network to decouple the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of the user interaction object.

Optionally, in an example of the above aspect, determining the total loss function based on the user interest prediction result and the user conformity prediction result may include: determining a user interest loss term based on the user interest prediction result; determining a user conformity loss term based on the user conformity prediction result; and determining the total loss function based on the user interest loss term and the user conformity loss term.

Optionally, in an example of the above aspect, the target object has an object popularity determined based on the number of interactions in the object interaction history behavior, the user interest loss item is determined based on the user interest prediction result and the object popularity of the target object, and the user conformity loss item is determined based on the user conformity prediction result and the object popularity of the target object.

Optionally, in an example of the above aspect, providing the user features, the object features of the user interaction object, and the object features of the target object to the embedding layer of the object recommendation model to obtain the user feature embedding representation of the user features, the object feature embedding representation of the user interaction object, and the object feature embedding representation of the target object may include: providing the user features, the object features of the user interaction object, the object features of the target object, and the object type of the user interaction object to the embedding layer of the object recommendation model to obtain the user feature embedding representation of the user features, the object feature embedding representation of the user interaction object, the object feature embedding representation of the target object, and the object type embedding representation of the user interaction object. Decoupling the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of the user interaction object via the decoupling layer of the object recommendation model may include: decoupling the user interest embedding representation, the user conformity embedding representation, and the user risk preference embedding representation corresponding to the user interest, the user conformity psychology, and the user risk preference respectively from the object feature embedding representation of the user interaction object via the decoupling layer of the object recommendation model. Providing the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation and the target object representation may include: providing the obtained fusion result, the object feature embedding representation of the target object, and the object type embedding representation of the user interaction object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation, the target object representation and the user risk preference representation. Determining the total loss function based on the user interest prediction result and the user conformity prediction result may include: determining the total loss function based on the user interest prediction result, the user conformity prediction result, the risk preference embedding representation and the user risk preference representation.

Optionally, in an example of the above aspect, determining the total loss function based on the user interest prediction result, the user conformity prediction result, the risk preference embedding representation and the user risk preference representation may include: determining a user interest loss term based on the user interest prediction result; determining a user conformity loss term based on the user conformity prediction result; determining a user risk preference loss term based on the risk preference embedding representation and the user risk preference representation; and determining the total loss function based on the user interest loss term, the user conformity loss term and the user risk preference loss term.

Optionally, in an example of the above aspect, the user risk preference loss term has a weighted value, and the weighted value is used to define the contribution of the user risk preference loss term to the total loss function. Determining the total loss function according to the user interest loss term, the user conformity loss term, and the user risk preference loss term may include: determining the total loss function according to the user interest loss term, the user conformity loss term, and the weighted user risk preference loss term.

Optionally, in an example of the above aspect, the object knowledge graph is created based on the relationship between the recommended type objects and the associated type objects.

According to another aspect of an embodiment of the present specification, a method for determining a recommended object based on an object recommendation model is provided, comprising: providing user features, object features of user interaction objects, and object features of a target object to an embedding layer of the object recommendation model, respectively, to obtain a user feature embedding representation of the user features, an object feature embedding representation of the user interaction objects, and an object feature embedding representation of the target object; decoupling a user interest embedding representation and a user conformity embedding representation, which correspond to user interests and user herd mentality, respectively, from the object feature embedding representation of the user interaction objects via a decoupling layer of the object recommendation model; fusing the user feature embedding representation with the user interest embedding representation and the user herd embedding representation, respectively, and providing the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model, respectively, to obtain a user interest representation, a user herd representation, and a target object representation; and determining whether to recommend the target object to the user based on the user interest representation, the user herd representation, and the target object representation.

Optionally, in an example of the above aspect, determining whether to recommend the target object to the user based on the user interest representation, the user conformity representation and the target object representation may include: predicting a user interest prediction result and a user conformity prediction result based on the user interest representation and the target object representation and based on the user conformity representation and the target object representation, respectively; and determining whether to recommend the target object to the user based on the user interest prediction result and the user conformity prediction result.

Optionally, in an example of the above aspect, the method may also include: providing the object feature embedding representations of the target object and the user interaction object to the graph neural network of the object recommendation model respectively to perform graph enhancement processing based on the object knowledge graph, so as to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing.

Optionally, in an example of the above aspect, the object recommendation model is trained according to the method described above.

According to another aspect of an embodiment of the present specification, there is provided an apparatus for training an object recommendation model, wherein the object recommendation model is used to determine whether to recommend a target object to a user, the apparatus comprising: a first representation generation unit, which provides user features, object features of user interaction objects, and object features of target objects to an embedding layer of the object recommendation model, respectively, to obtain a user feature embedding representation of the user features, an object feature embedding representation of the user interaction objects, and an object feature embedding representation of the target object; an embedding representation decoupling unit, which decouples a user interest embedding representation and a user conformity embedding representation, which correspond to user interests and user conformity psychology, respectively, from the object feature embedding representation of the user interaction objects via the decoupling layer of the object recommendation model; a second representation generation unit, which fuses the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, respectively, and combines the obtained fusion result with the target object embedding representation. The object feature embedding representation of the image is respectively provided to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation and the target object representation; the target object prediction unit predicts the user interest prediction result based on the user interest representation and the target object representation, and predicts the user conformity prediction result based on the user conformity representation and the target object representation; the loss function determination unit determines the total loss function based on the user interest prediction result and the user conformity prediction result in response to the training end condition not being met; and the model adjustment unit adjusts the model parameters of the object recommendation model according to the total loss function, wherein the first representation generation unit, the graph enhancement processing unit, the embedding representation decoupling unit, the second representation generation unit, the third representation generation unit, the target object prediction unit, the loss function determination unit and the model adjustment unit perform operations in a loop until the training end condition is met.

According to another aspect of an embodiment of the present specification, there is provided an apparatus for determining a recommended object based on an object recommendation model, comprising: a first representation generation unit, which provides user features, object features of user interaction objects, and object features of target objects to an embedding layer of the object recommendation model, respectively, to obtain a user feature embedding representation of the user features, an object feature embedding representation of the user interaction objects, and an object feature embedding representation of the target object; an embedding representation decoupling unit, which decouples a user interest embedding representation and a user conformity embedding representation corresponding to user interests and user conformity, respectively, from the object feature embedding representation of the user interaction objects via the decoupling layer of the object recommendation model; a second representation generation unit, which fuses the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, respectively, and provides the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model, respectively, to obtain a user interest representation, a user conformity representation, and a target object representation; and a recommendation decision unit, which determines whether to recommend the target object to the user based on the user interest representation, the user conformity representation, and the target object representation.

Optionally, in an example of the above aspect, the device may also include: a graph enhancement processing unit, which provides the object feature embedding representations of the target object and the user interaction object to the graph neural network of the object recommendation model respectively to perform graph enhancement processing based on the object knowledge graph to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing.

According to another aspect of an embodiment of the present specification, there is provided an object recommendation system, comprising: an interactive object acquisition device for acquiring a user interactive object; a device for training an object recommendation model as described above; and a device for determining a recommended object based on the object recommendation model as described above.

According to another aspect of an embodiment of the present specification, there is provided an apparatus for training an object recommendation model, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the method for training an object recommendation model as described above.

According to another aspect of an embodiment of the present specification, there is provided an apparatus for determining a recommended object based on an object recommendation model, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the method for determining a recommended object based on an object recommendation model as described above.

BRIEF DESCRIPTION OF THE DRAWINGS

A further understanding of the nature and advantages of the content of this specification may be achieved by referring to the following drawings. In the accompanying drawings, similar components or features may have the same reference numerals.

FIG1 shows a block diagram of an object recommendation system based on an object recommendation model according to an embodiment of the present specification.

FIG. 2 shows an example schematic diagram of an object recommendation model according to an embodiment of the present specification.

FIG. 3 shows an example flow chart of a method for training an object recommendation model according to an embodiment of the present specification.

FIG. 4 shows an example flow chart of a total loss function determination process according to an embodiment of the present specification.

FIG. 5 shows an example schematic diagram of an object recommendation model training process according to an embodiment of the present specification.

FIG. 6 shows an example flow chart of a method for determining a recommended object based on an object recommendation model according to an embodiment of the present specification.

FIG. 7 shows an example schematic diagram of an object recommendation process according to an embodiment of the present specification.

FIG. 8 shows an example block diagram of an object recommendation model training device according to an embodiment of this specification.

FIG. 9 shows an example block diagram of a loss function determination unit according to an embodiment of the present specification.

FIG. 10 shows an exemplary block diagram of an object recommendation device according to an embodiment of the present specification.

FIG. 11 shows an example schematic diagram of an object recommendation model training device implemented based on a computer system according to an embodiment of the present specification.

FIG. 12 shows an example schematic diagram of an object recommendation device implemented based on a computer system according to an embodiment of the present specification.

DETAILED DESCRIPTION

The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and is not a limitation of the scope of protection, applicability or examples set forth in the claims. The functions and arrangements of the elements discussed can be changed without departing from the scope of protection of the contents of this specification. Various examples can omit, replace or add various processes or components as needed. For example, the described method can be performed in an order different from the described order, and various steps can be added, omitted or combined. In addition, the features described relative to some examples can also be combined in other examples.

As used herein, the term "including" and its variations represent open terms, meaning "including but not limited to". The term "based on" means "based at least in part on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other definitions may be included below, whether explicit or implicit. Unless the context clearly indicates otherwise, the definition of a term is consistent throughout the specification.

When recommending objects to users, the recommended objects are usually determined based on the object recommendation model and recommended to the users. Object recommendation models can usually be modeled based on sequential behavior or based on decoupled learning. However, modeling schemes based on sequential behavior usually only learn embedded representations dominated by user interests based on user historical behavior, without distinguishing between potential behavioral intentions. Modeling schemes based on decoupled learning also find it difficult to explicitly include user intentions such as user conformity and user risk preference when recommending objects.

In some object recommendation application scenarios, such as commodity intelligent recommendation scenarios and fund intelligent recommendation scenarios, when users select objects, in addition to the user's interest in the selected object, whether the selected object is selected by many users and whether the risk of the object matches the user's risk preference will affect the user's object selection. Therefore, when modeling the object recommendation model, in addition to considering user interests, factors such as user conformity and user risk preference need to be considered. Therefore, the above-mentioned modeling scheme based on sequential behavior and the modeling scheme based on decoupled learning are not suitable for application in the above-mentioned object recommendation scenarios that need to consider user conformity and user risk preference.

In view of this, the embodiments of the present specification provide an object recommendation model training scheme and an object recommendation scheme based on the object recommendation model. Using the object recommendation model training scheme, the user feature embedding representation of the user feature, the object feature embedding representation of the target object and the user interaction object are obtained by providing the user feature, the target object and the user interaction object's object features to the embedding layer of the object recommendation model, and the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity are respectively decoupled from the object feature embedding representation of the user interaction object through the decoupling layer of the object recommendation model; then the decoupled user interest embedding representation and the user conformity embedding representation are introduced into the object recommendation model training, thereby modeling the object recommendation model based on the user interest and the user conformity, thereby improving the user understanding effect of the object recommendation model, so that the trained object recommendation model is suitable for the above-mentioned object recommendation scenario that needs to consider the user interest and the user conformity.

The following describes an object recommendation model training method and an object recommendation model training device, a recommended object determination method and a recommended object determination device, and an object recommendation system according to an embodiment of the present specification with reference to the accompanying drawings.

FIG. 1 shows a block diagram of an object recommendation system 100 based on an object recommendation model according to an embodiment of the present specification.

As shown in FIG1 , the object recommendation system 100 includes an interactive object acquisition device 110, an object recommendation model training device 120, an object recommendation model storage device 130, and an object recommendation device 140. The interactive object acquisition device 110, the object recommendation model training device 120, the object recommendation model storage device 130, and the object recommendation device 140 can communicate with each other through a network 150. In some embodiments, the network 150 can be any one or more of a wired network or a wireless network. Examples of the network 150 may include, but are not limited to, a cable network, an optical fiber network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network (ZigBee), a near field communication (NFC), an in-device bus, an in-device line, etc., or any combination thereof. In some embodiments, some or all of the interactive object acquisition device 110, the object recommendation model training device 120, the object recommendation model storage device 130, and the object recommendation device 140 can communicate directly without the network 150.

The interactive object acquisition device 110 is used to acquire user interactive objects. User interactive objects refer to objects that have historical interactive behaviors with users, that is, objects that users have previously interacted with. Examples of the term "interactive behavior" may include, but are not limited to, interactive behaviors such as users clicking on objects, users viewing objects, users browsing objects, users purchasing objects, users consulting objects, and users operating objects. The term "object" may refer to an object suitable for recommendation to users. Examples of objects may include, but are not limited to, commodities, fund products, video materials such as fitness programs/learning programs, etc. Each object may have an object type and an object feature. The object type is used to identify the type attribute of the object. For example, when the object is a commodity, the object type may include ordinary commodities, high-end commodities, luxury commodities, fragile commodities, high-renewal commodities, etc. When the object is a fund product, the object type may include stock funds, bond funds, futures funds, stock/bond hybrid funds, etc. The object type may have object risk information for identifying object risks. For example, stock funds, bond funds, futures funds, and stock/bond hybrid funds can identify different fund product risks. Object features may refer to features used to reflect the characteristics of an object, and the object features may distinguish the corresponding object from other objects. When the object is a fund product, the object features may include a fund company, a fund manager, and the like.

In some embodiments, the interactive object acquisition device 110 can collect user interactive objects by monitoring the user's historical interactive behavior. In some embodiments, the user interactive objects can be collected via a specific object collection device. The interactive object device 110 can obtain user interactive objects from a specific object collection device via wireless communication or wired communication.

The user interaction object acquired by the interaction object acquisition device 110 can be provided to the object recommendation model training device 120. The object recommendation model training device 120 can use the user interaction object and the target object to train the object recommendation model. When performing object recommendation model training, the target object can be one or more pre-selected object samples for model training, and the object sample has a behavior label for a specific user (e.g., user u). For example, if there is an interaction behavior between user u and the target object, the label value of the target object is 1. If there is no interaction behavior between user u and the target object, the label value of the target object is 0. The object recommendation model training process of the object recommendation model training device 120 will be described in detail below with reference to the accompanying drawings.

The object recommendation device 140 may determine whether to recommend the target object based on the trained object recommendation model. In response to determining to recommend the target object, the object recommendation device 140 may recommend the target object to the user. In some embodiments, the object recommendation model trained by the object recommendation model training device 120 may be stored in the object recommendation model storage device 130. In this case, when the object recommendation device 140 makes an object recommendation, the object recommendation model may be obtained from the object recommendation model storage device 130 to make an object recommendation, or the object recommendation model storage device 130 may be communicated with to provide the information required by the object recommendation model to the object recommendation model in the object recommendation model storage device 130 to make an object recommendation. In some embodiments, the object recommendation model trained by the object recommendation model training device 120 may be stored in the object recommendation device 140.

Fig. 2 shows an example schematic diagram of an object recommendation model 200 according to an embodiment of the present specification. As shown in Fig. 2, the object recommendation model 200 may include an embedding layer 210, a graph neural network layer 220, a decoupling layer 230, a cascade layer 240, a fusion layer 250, a representation layer 260, and a prediction layer 270.

The embedding layer 210 is used to map the input user features, the object features of the target object, the object features of the user interaction object, and the object type of the user interaction object to the corresponding embedded representation in the representation space. Here, the user interaction object may include one or more user interaction objects. For example, the embedding layer 210 can map the input user features to the user feature embedded representation in the representation space. The embedding layer 210 can map the object features of the input target object to the object feature embedded representation of the target object in the representation space. The embedding layer 210 can map the object features of the input user interaction object to the object feature embedded representation of the user interaction object in the representation space. The embedding layer 210 can map the object type of the input user interaction object to the object type embedded representation of the user interaction object in the representation space. In some embodiments, the embedding layer 210 can have multiple embedding layers, such as embedding layer 1 to embedding layer N. Each embedding layer is responsible for the embedding representation mapping process of a feature or information. For example, an embedding layer can be assigned to each user feature and the object type of each user interaction object to perform the embedding representation mapping process.

The graph neural network layer 220 is used to perform graph enhancement processing on the object feature embedding representation of the target object and the user interaction object based on the object knowledge graph, thereby obtaining the object feature embedding representation of the target object and the user interaction object after the graph enhancement processing. The graph neural network layer 220 may include one or more graph neural networks. For example, a graph neural network can be assigned to each object, and the graph neural network is used to perform graph enhancement processing on the object feature embedding representation of the object based on the object knowledge graph, thereby obtaining the object feature embedding representation of the object after the graph enhancement processing. The graph neural network can be implemented using any graph neural network in the art. Examples of graph neural networks may include, for example, but are not limited to, graph convolution networks (Graph Convolution Networks, GCN), graph attention networks (Graph Attention Networks), graph autoencoders (Graph Autoencoders), graph generative networks (Graph Generative Networks) and graph spatial-temporal networks (Graph Spatial-temporal Networks), etc.

In some embodiments, an object knowledge graph can be created based on the relationship between a recommendation type object and an associated type object. A recommendation type object refers to an object of the same type as a target object. An associated type object refers to an object that has an associated relationship with a recommendation type object. For example, when the object is a fund product, the associated type object may include a fund manager, a fund company, the industry to which the fund belongs, heavily weighted stocks, a market index, and the like. When creating a knowledge graph, various nodes and edge relationships in the fund knowledge graph can be created based on the five relationships of fund product->fund manager, fund product->fund company, fund product->fund industry, fund product->heavy weighted stocks, and fund product->market index, thereby constructing the association relationship between all fund products. Using a graph neural network to perform graph convolution operations based on an object knowledge graph, the object knowledge graph can be used to introduce similar objects for a given object to enhance the graph representation of the given object.

The object feature embedding representation of the user interaction object after graph enhancement processing by the graph neural network layer 220 is provided to the decoupling layer 230. The decoupling layer 230 decouples the user interest embedding representation, user conformity embedding representation and user risk preference embedding representation corresponding to the user interest, user conformity and user risk preference from the object feature embedding representation of the user interaction object. In some embodiments, the decoupling layer 230 can be implemented as a self-attention network.

Optionally, when the user interaction object includes multiple user interaction objects, the object recommendation model 200 may further include a cascade layer 240. The feature embedding representation components corresponding to user interests, user conformity, and user risk preferences decoupled by the decoupling layer 230 are provided to the cascade layer 240 for cascading, thereby obtaining user interest embedding representation, user conformity embedding representation, and user risk preference embedding representation. It should be noted that, in some embodiments, the cascade layer 240 can be integrated in the decoupling layer 230. It should be noted that the term "cascade" can be used interchangeably with splicing.

The user interest embedding representation and the user conformity embedding representation obtained by cascading the decoupling layer 230/the cascading layer 240 and the user feature embedding representation output by the embedding layer 210 are provided to the fusion layer 250. The fusion layer 250 fuses the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, respectively, and provides the obtained fusion results to the representation layer 260. Here, the purpose of the fusion layer 250 is to integrate the user features into the user interest embedding representation and the user conformity embedding representation, so that the fusion result can reflect the user interest embedding representation and the user conformity embedding representation combined with the user features. In some embodiments, the fusion layer 250 can be implemented as a cascade layer or other applicable fusion technology.

The representation layer 260 is used to obtain the user interest representation based on the fusion result of the user feature embedding representation and the user interest embedding representation, and to obtain the user conformity representation based on the fusion result of the user feature embedding representation and the user conformity embedding representation. In addition, the representation layer 260 also receives the object type embedding representation of the user interaction object provided by the embedding layer 210, and obtains the user risk preference representation based on the object type embedding representation of the user interaction object. In addition, the representation layer 260 also receives the object feature embedding representation of the target object provided by the graph neural network layer 220, and obtains the target object representation based on the object feature embedding representation of the target object. In some embodiments, the representation layer 260 can be implemented as, for example, a feedforward network (FFN) or a feedforward neural network (FFNN).

The user interest representation, user conformity representation, and target object representation obtained by the representation layer 260 are provided to the prediction layer 270. The prediction layer 270 predicts a user interest prediction result based on the user interest representation and the target object representation, and predicts a user conformity prediction result based on the user conformity representation and the target object representation, and predicts whether the target object is a recommended object based on the user interest prediction result and the user conformity prediction result.

It should be noted that the above example is only an exemplary embodiment of the present disclosure, and the included features are not all necessary features of the present disclosure. As an example, although the graph neural network layer 220 is included in the above example, this feature is only exemplary and not necessary. The target object and the user interaction object may not be represented in the form of a graph, so the object recommendation model may not include the graph neural network layer 220. In addition, similarly, although the object type and user risk preference representation of the user interaction object are included in the above example, these features are also only exemplary and are used to better extract the user conformity embedding representation, but are not necessary. When modeling the object recommendation model, only the user interest embedding representation and the user conformity embedding representation can be used without using the user risk preference representation, and accordingly, the object type of the user interaction object may not be used.

Fig. 3 shows an example flow chart of a method 300 for training the object recommendation model 200 shown in Fig. 2 according to an embodiment of the present specification. The object recommendation model training process shown in Fig. 3 is a cyclic process.

As shown in Figure 3, in each loop process, at 310, the user features, the object features of the target object, the object features of the user interaction object, and the object type of the user interaction object are respectively provided to the embedding layer of the object recommendation model to obtain the user feature embedding representation of the user features, the object feature embedding representation of the target object, the object feature embedding representation of the user interaction object, and the object type embedding representation of the user interaction object. In this specification, the user interaction object and the target object belong to the same type of object, and the object type has object risk information. For example, when the object is a fund product, the user interaction object is a user interaction fund product, and the target object is a target fund product. The object type is a fund product type, for example, a stock fund, a bond fund, a futures fund, etc. Through the fund product type, the risk level of the fund product can be reflected, for example, low risk, medium risk, high risk, etc.

At 320, the object feature embedding representations of the target object and the user interaction object are respectively provided to the graph neural network of the object recommendation model to perform graph enhancement processing based on the object knowledge graph, so as to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing.

At 330, via the decoupling layer of the object recommendation model, the user interest embedding representation, the user conformity embedding representation and the user risk preference embedding representation corresponding to the user interest, the user conformity mentality and the user risk preference are decoupled from the object feature embedding representation after graph enhancement processing of the user interaction object.

In some embodiments, the decoupling layer of the object recommendation model can be used to decouple the user interest embedding representation, user conformity embedding representation and user risk preference embedding representation corresponding to user interests, user conformity and user risk preferences respectively from the object feature embedding representation of the user interaction object after graph enhancement based on an unsupervised mechanism.

In some embodiments, the user interaction object may include multiple user interaction objects. In this case, the feature embedding representation components corresponding to user interests, user conformity and user risk preferences can be decoupled from the object feature embedding representations of each user interaction object after graph enhancement through the decoupling layer of the object recommendation model and cascaded to obtain the user interest embedding representation, the user conformity embedding representation and the user risk preference embedding representation.

For example, in some embodiments, the decoupling layer can be implemented as a self-attention network. In this case, the object feature embedding representations of each user interaction object after graph enhancement processing can be provided to the self-attention network respectively, so as to decouple the feature embedding representation components corresponding to the user interest, the user herd mentality and the user risk preference from the object feature embedding representations of each user interaction object after graph enhancement processing.

针对该对象行为序列表征

可以使用自注意力机制来抽取用户兴趣，用户从众心理和用户风险偏好的向量集合{ω^I,ω^C,ω^R}进行下述操作，由此得到解耦化的对象特征嵌入表征：For example, suppose a graph neural network is used to perform graph convolution operations on the object features of a given object based on the object knowledge graph to obtain the object feature embedding representation set after graph enhancement processing as H ^(L) ∈R ^|ε|×d , where |ε| is the number of objects and d is the size of the graph representation. Given the historical interaction sequence of user u as S = {f ₁ ,…,f _|S| }, the object behavior sequence representation can be retrieved from the object feature embedding representation set H ^(L) after graph enhancement processing

Characterize the behavior sequence of the object

The self-attention mechanism can be used to extract the vector set {ω ^I ,ω ^C ,ω ^R } of user interests, user conformity and user risk preference, and perform the following operations to obtain the decoupled object feature embedding representation:

Among them, σ(·) is a nonlinear function and f(·) is a Softmax function.

Through the above self-attention operation, we can obtain the multi-granularity decoupled object feature embedding representation for user u based on the unsupervised mechanism:

At 340, the user feature embedding representation is fused with the user interest embedding representation and the user conformity embedding representation respectively, and the two fusion results are provided to the representation layer of the object recommendation model to obtain the user interest representation and the user conformity representation.

At 350 , the object feature embedding representation of the target object is provided to the representation layer of the object recommendation model to obtain the target object representation, and the object type embedding representation of the user interaction object is provided to the representation layer of the object recommendation model to obtain the user risk preference representation.

At 360, a prediction layer of the object recommendation model predicts a user interest prediction result based on the user interest representation and the target object representation, and predicts a user conformity prediction result based on the user conformity representation and the target object representation. The user interest prediction result refers to a prediction result about the user interest, and the user conformity prediction result is a prediction result about the user conformity. That is, the prediction result about the user interest is obtained by calculating the similarity between the user interest representation and the target object representation, and the prediction result about the user conformity is obtained by calculating the similarity between the user conformity representation and the target object representation.

可以基于

来确定出，其中，σ(·)为sigmoid函数，FFN^I(·)为与用户兴趣表征对应的前向传播网络，

为用户特征嵌入表征，

为用户兴趣嵌入表征，||为级联操作，(·)^T为转置操作，以及x_f为目标对象表征。In some embodiments, the user interest prediction result

Can be based on

To determine, where σ(·) is the sigmoid function, ^FFNI (·) is the forward propagation network corresponding to the user interest representation,

Embedding representation for user features,

is the user interest embedding representation, || is the concatenation operation, (·) ^T is the transposition operation, and x _f is the target object representation.

可以基于

来确定出，其中，σ(·)为sigmoid函数，FFN^C(·)为与用户从众表征对应的前向传播网络，

为用户特征嵌入表征，

为用户从众嵌入表征，||为级联操作，(·)^T为转置操作，以及x_f为目标对象表征。In some embodiments, the user follows the crowd to predict the results

Can be based on

To determine, where σ(·) is the sigmoid function, FFN ^C (·) is the forward propagation network corresponding to the user's conformity representation,

Embedding representation for user features,

is the user’s crowd embedding representation, || is the concatenation operation, (·) ^T is the transposition operation, and x _f is the target object representation.

At 370, it is determined whether the model training process of the object recommendation model is completed. For example, it is possible to determine whether the model training process is completed by determining whether a model training completion condition is met. Examples of model training completion conditions may include, but are not limited to: reaching a preset number of training rounds, or the trained model accuracy reaching a predetermined accuracy, etc. When the model training completion condition is that the trained model accuracy reaches a predetermined accuracy, the prediction result for the target object may be determined based on the user interest prediction result and the user conformity prediction result, and the trained model accuracy may be determined based on the prediction result for the target object.

If the training end condition is met, the model training process ends. If the training end condition is not met, at 380, a total loss function is determined based on the user interest prediction result, the user conformity prediction result, the risk preference embedding representation, and the user risk preference representation.

FIG. 4 shows an example flow chart of a total loss function determination process 400 according to an embodiment of the present specification.

As shown in FIG. 4 , at 410 , a user interest loss term is determined based on the user interest prediction result, and at 420 , a user conformity loss term is determined based on the user conformity prediction result.

In some embodiments, the target object may have an object popularity determined based on the number of interactions in the object's interaction history behavior.

In some embodiments, the popularity of the target object f may be defined as follows:

Among them, γ _f is the popularity of the target object f, C _f is the number of interactions of the target object f in the historical data, C _max is the maximum number of interactions of each object in the object set in the historical data, and C _min is the minimum number of interactions of each object in the object set in the historical data.

In this case, the user interest loss term may be determined based on the user interest prediction result and the object popularity of the target object, and the user conformity loss term may be determined based on the user conformity prediction result and the object popularity of the target object.

其中，

为用户兴趣预测结果，

为用户u对目标对象f的行为标签，以及CE(·)为交叉熵损失函数。For example, the user interest loss term ^LI can be expressed as:

in,

Predict results for user interests,

is the behavior label of user u on target object f, and CE(·) is the cross entropy loss function.

其中，

为用户从众预测结果，

为用户u对目标对象f的行为标签，以及CE(·)为交叉熵损失函数。The user conformity loss term ^LC can be expressed as:

in,

Predict results for users to follow the crowd,

is the behavior label of user u on target object f, and CE(·) is the cross entropy loss function.

At 430 , a user risk preference loss term is determined based on the risk preference embedding representation and the user risk preference representation.

可以从中抽取出特定表示并将其作为风险偏好的自监督信息，该特定表示可以定义如下：Given that object types can well characterize the risk level of an object, useful prior information can be extracted from the object type sequence in historical behavior to model the user's risk preference. For example, assuming that the historical fund type sequence of a given user u is

A specific representation can be extracted from it and used as self-supervisory information for risk preference. The specific representation can be defined as follows:

Among them, Φ(·) is the Embedding operator, g(·) is the Pooling function (for example, it can be defined as the Concact operation), and FFN(·) is the forward propagation network.

Based on the above definition, the user risk preference loss term ^LR can be defined as the following contrast loss function as a self-supervised loss function:

中采样得到，以及τ为温度参数。在一些实施例中，分布

被实现为均匀分布。在其它实施例中，分布

被也可以采用其他的有偏分布。温度参数τ可以用来控制施加在不同负样本上的惩罚强度。Among them, B is a specific batch of data, and negative samples can be based on the distribution of noise

In some embodiments, the distribution

is implemented as a uniform distribution. In other embodiments, the distribution

Other biased distributions can also be used. The temperature parameter τ can be used to control the intensity of the penalty imposed on different negative samples.

At 440 , a total loss function is determined based on the user interest loss term, the user conformity loss term, and the user risk preference loss term.

In some embodiments, the user interest loss term, the user conformity loss term, and the user risk preference loss term may be used to obtain a total loss function, that is, L= ^LI + ^LC + ^LR . In some embodiments, the user risk preference loss term ^LR may have a weighted value ε, which is used to define the contribution of the user risk preference loss term to the total loss function. In this case, the total loss function may be determined based on the user interest loss term, the user conformity loss term, and the weighted user risk preference loss term. For example, L= ^LI + ^LC +ε· ^LR . ε is a constant greater than 0. The constant ε is used to control the contribution of the risk preference term ^LR to the total loss function, while the contribution of the user interest term ^LI and the user conformity psychology term ^LC to the total loss function may be adaptively adjusted by the object popularity parameter _γf .

Returning to FIG. 3 , after the total loss function is determined as above, at 390 , the model parameters of the object recommendation model are adjusted according to the total loss function. For example, the model parameters of one or more layers of the embedding layer, the graph neural network layer, the representation layer, and the prediction layer in the object recommendation model are adjusted. Then, the process returns to 310 to execute the next loop process. This cycle is repeated until the model training process is completed.

It should be noted that FIG. 3 and FIG. 4 only illustrate an exemplary embodiment of the training process of the object recommendation model 200 shown in FIG. 2. As described above with respect to FIG. 2, some features of the object recommendation model 200 of FIG. 2 are not essential to the present disclosure. Similarly, not all steps of the training process shown in FIG. 3 and FIG. 4 are essential to the present disclosure.

As an example, in the example of FIG. 4 , the object recommendation model is modeled using the user interest embedding representation, the user conformity embedding representation, and the user risk preference embedding representation. In some embodiments, only the user interest embedding representation and the user conformity embedding representation can be used to model the object recommendation model. In this case, the operation related to the user risk preference embedding representation can be modified from the training process illustrated in FIGS. 3 and 4 . For example, in 310 , the object type of the user interaction object does not need to be provided to the embedding layer of the object recommendation model to obtain the object type embedding representation of the user interaction object. Thus, in 330 , the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology are decoupled from the object feature embedding representation of the user interaction object after the graph enhancement processing via the decoupling layer of the object recommendation model. In 350 , only the object feature embedding representation of the target object is provided to the representation layer of the object recommendation model to obtain the target object representation. In addition, in 380 , the total loss function can be determined based on the user interest prediction result and the user conformity prediction result. Specifically, the user interest loss term is determined based on the user interest prediction result, and the user conformity loss term is determined based on the user conformity prediction result. Then, the total loss function is determined based on the user interest loss term and the user conformity loss term.

In some embodiments, the object recommendation model may not include a graph neural network, so that when training the object recommendation model, the object feature embedding representation of the target object and the user interaction object does not perform graph enhancement processing based on the object knowledge graph. In some embodiments, only the user interest embedding representation and the user conformity embedding representation may be used to model the object recommendation model, and the object feature embedding representation of the target object and the user interaction object does not perform graph enhancement processing based on the object knowledge graph.

The object recommendation model training process according to the embodiment of the present specification is described above with reference to Figures 1 to 4. The object recommendation model training process is described below using commodity recommendation as an example.

Fig. 5 shows an example schematic diagram of an object recommendation model training process according to an embodiment of the present specification. In the example of Fig. 5, the object is a commodity, and the interactive objects include interactive commodity 1 to interactive commodity n.

交互商品1的对象特征被提供给嵌入层得到交互商品1的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理，由此得到交互商品1的经过图增强处理后的对象特征嵌入表征。交互商品2的对象特征被提供给嵌入层得到交互商品2的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理，由此得到交互商品2的经过图增强处理后的对象特征嵌入表征，以及交互商品n的对象特征被提供给嵌入层得到交互商品n的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到交互商品n的经过图增强处理后的对象特征嵌入表征。目标商品的对象特征被提供给嵌入层得到交互商品的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到目标商品的经过图增强处理后的对象特征嵌入表征x_f。The user features are provided to the embedding layer to obtain the user feature embedding representation

The object features of interactive product 1 are provided to the embedding layer to obtain the object feature embedding representation of interactive product 1, and the obtained object feature embedding representation is subjected to graph enhancement processing using the product knowledge graph through the graph neural network, thereby obtaining the object feature embedding representation of interactive product 1 after graph enhancement processing. The object features of interactive product 2 are provided to the embedding layer to obtain the object feature embedding representation of interactive product 2, and the obtained object feature embedding representation is subjected to graph enhancement processing using the product knowledge graph through the graph neural network, thereby obtaining the object feature embedding representation of interactive product 2 after graph enhancement processing, and the object features of interactive product n are provided to the embedding layer to obtain the object feature embedding representation of interactive product n, and the obtained object feature embedding representation is subjected to graph enhancement processing using the product knowledge graph through the graph neural network to obtain the object feature embedding representation of interactive product n after graph enhancement processing. The object features of the target product are provided to the embedding layer to obtain the object feature embedding representation of the interactive product, and the obtained object feature embedding representation is subjected to graph enhancement processing using the product knowledge graph through the graph neural network to obtain the object feature embedding representation x _f of the target product after graph enhancement processing.

目标对象的经过图增强处理后的对象特征嵌入表征被提供给对象推荐模型的表征层得到目标对象表征。The object types of interactive product 1, interactive product 2 to interactive product n are respectively provided to the embedding layer to obtain the object type embedding representation of interactive product 1, the object type embedding representation of interactive product 2, and the object type embedding representation of interactive product n. Subsequently, the object type embedding representation of interactive product 1, the object type embedding representation of interactive product 2, and the object type embedding representation of interactive product n are provided to the feed-forward network (FFN) to obtain the user risk preference representation

The object feature embedding representation of the target object after graph enhancement processing is provided to the representation layer of the object recommendation model to obtain the target object representation.

将从交互商品1到交互商品n的经过图增强处理后的对象特征嵌入表征中解耦出的与用户从众心理对应的对象特征嵌入表征分量级联，得到用户从众嵌入表征

将从交互商品1到交互商品n的经过图增强处理后的对象特征嵌入表征中解耦出的与用户风险偏好对应的对象特征嵌入表征分量级联，得到用户风险偏好嵌入表征

The object feature embedding representation components corresponding to user interests, user herd mentality, and user risk preference are decoupled from the object feature embedding representation of interactive product 1 after graph enhancement. The object feature embedding representation components corresponding to user interests, user herd mentality, and user risk preference are decoupled from the object feature embedding representation of interactive product 2 after graph enhancement. The object feature embedding representation components corresponding to user interests, user herd mentality, and user risk preference are decoupled from the object feature embedding representation of interactive product n after graph enhancement. Then, the object feature embedding representation components corresponding to user interests decoupled from the object feature embedding representations of interactive product 1 to interactive product n after graph enhancement are cascaded to obtain the user interest embedding representation

The object feature embedding representation components corresponding to the user's herd mentality that are decoupled from the object feature embedding representation after graph enhancement from interactive product 1 to interactive product n are cascaded to obtain the user herd embedding representation

The object feature embedding representation components corresponding to the user risk preference decoupled from the object feature embedding representation after graph enhancement from interactive product 1 to interactive product n are cascaded to obtain the user risk preference embedding representation

与用户特征嵌入表征融合(级联)，并将融合结果提供给前馈网络得到用户兴趣表征。将用户从众嵌入表征

与用户特征嵌入表征融合，并将融合结果提供给前馈网络得到用户从众表征。然后，基于用户兴趣表征和目标对象表征得到用户兴趣预测结果

以及基于用户从众表征和目标对象表征得到用户从众预测结果

Then, user interests are embedded into representation

Fuse (cascade) with the user feature embedding representation, and provide the fusion result to the feedforward network to obtain the user interest representation.

The user feature embedding representation is fused and the fusion result is provided to the feedforward network to obtain the user conformity representation. Then, the user interest prediction result is obtained based on the user interest representation and the target object representation.

And the user conformity prediction result is obtained based on the user conformity representation and the target object representation

和用户从众预测结果

后，基于用户兴趣预测结果

和用户u对目标商品f的行为标签

确定用户兴趣损失项L^I，以及基于用户从众预测结果

和用户u对目标商品f的行为标签

确定用户从众损失项L^C。此外，还基于用户风险偏好嵌入表征

和用户风险偏好表征

确定用户风险偏好损失项L^R。After determining the user interest prediction results

and user herd prediction results

After that, based on the user interest prediction results

and the behavior label of user u on the target product f

Determine the user interest loss term ^LI and the prediction result based on user conformity

and user u's behavior label for target product f

Determine the user's herd loss term ^LC . In addition, embedding representation based on user risk preference

and user risk preference characterization

Determine the user's risk preference loss term ^LR .

Subsequently, a total loss function L is determined based on the user interest loss term ^LI , the user conformity loss term ^LC, and the user risk preference loss term ^LR , and the model parameters of the object recommendation model are adjusted according to the total loss function L.

By utilizing the above-mentioned object recommendation model training scheme, user feature embedding representations of user features and object feature embedding representations of target objects and user interaction objects are obtained by providing user features, target objects, and user interaction objects to the embedding layer of the object recommendation model, and the user interest embedding representations and user conformity embedding representations corresponding to user interests and user conformity are respectively decoupled from the object feature embedding representations of user interaction objects through the decoupling layer of the object recommendation model; the decoupled user interest embedding representations and user conformity embedding representations are then introduced into the object recommendation model training, thereby modeling the object recommendation model based on user interests and user conformity, thereby improving the user understanding effect of the object recommendation model, and making the trained object recommendation model suitable for the above-mentioned object recommendation scenarios that require consideration of user interests and user conformity.

Utilizing the above-mentioned object recommendation model training scheme, by deploying a graph neural network in the object recommendation model, and using the graph neural network to perform graph enhancement processing on the object features of the target object and the user interaction object based on the object knowledge graph, the object feature embedding representation of the target object and the user interaction object after graph enhancement processing is obtained, so that the feature influence of the associated objects in the object knowledge graph can be introduced into the object feature embedding representation of the target object and the user interaction object, thereby improving the accuracy of the object feature embedding representation of the target object and the user interaction object.

Utilizing the above-mentioned object recommendation model training scheme, a graph neural network is deployed in the object recommendation model, and the graph neural network is used to perform graph enhancement processing on the object features of the target object and the user interaction object based on the object knowledge graph to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing, and then the user interest embedding representation, user conformity embedding representation and user risk preference embedding representation corresponding to the user interest, user herd mentality and user risk preference are respectively decoupled from the object feature embedding representation of the user interaction object; and then the decoupled user interest embedding representation, user herd embedding representation and user risk preference embedding representation are introduced into the object recommendation model training, thereby modeling the object recommendation model based on user interest, user herd mentality and user risk preference, thereby improving the user understanding effect of the object recommendation model, so that the trained object recommendation model is suitable for the above-mentioned object recommendation scenario that needs to consider the user herd mentality and user risk preference.

By using the above-mentioned object recommendation model training scheme, the model training accuracy of the object recommendation model can be further improved by setting the total loss function to be determined based on the user interest loss term, the user conformity loss term and the user risk preference loss term.

By using the above-mentioned object recommendation model training scheme and setting the object popularity, the contribution of the user interest loss term and the user conformity loss term to the total loss function can be adaptively adjusted, thereby improving the application scenario adaptability of the object recommendation model.

By using the above-mentioned object recommendation model training scheme and setting a weighted value for the risk preference loss item, the contribution of the risk preference item to the total loss function can be controlled, thereby further improving the application scenario adaptability of the object recommendation model.

After the object recommendation model is trained as above, it is possible to predict whether the target object is a recommended object based on the object recommendation model, thereby implementing object recommendation processing.

FIG. 6 shows an example flow chart of a method 600 for determining a recommended object based on an object recommendation model according to an embodiment of the present specification.

As shown in FIG. 6 , at 610 , user features, object features of the target object, and object features of the user interaction object are provided to the embedding layer of the object recommendation model to obtain user feature embedding representations of the user features and object feature embedding representations of the target object and the user interaction object.

At 620, the object feature embedding representations of the target object and the user interaction object are respectively provided to the graph neural network of the object recommendation model to perform graph enhancement processing based on the object knowledge graph, so as to obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing.

At 630 , a user interest embedding representation and a user conformity embedding representation corresponding to user interest and user conformity psychology are decoupled from the object feature embedding representation of the user interaction object after graph enhancement via the decoupling layer of the object recommendation model.

In some embodiments, the user interaction object includes multiple user interaction objects. In this case, the feature embedding representation components corresponding to user interests and user conformity can be decoupled from the object feature embedding representations of each user interaction object after graph enhancement through the decoupling layer of the object recommendation model and cascaded to obtain the user interest embedding representation and the user conformity embedding representation.

At 640, the user feature embedding representation is fused with the user interest embedding representation and the user conformity embedding representation respectively, and the obtained fusion results are provided to the representation layer of the object recommendation model to obtain the user interest representation and the user conformity representation.

At 650 , the object feature embedding representation of the target object is provided to a representation layer of the object recommendation model to obtain a target object representation.

Subsequently, based on the user interest representation, the user conformity representation and the target object representation, it is determined whether to recommend the target object to the user.

For example, in some embodiments, at 660 , a user interest prediction result is predicted based on the user interest representation and the target object representation, and a user conformity prediction result is predicted based on the user conformity representation and the target object representation.

其中，

为用户兴趣预测结果，以及

为用户从众预测结果。At 670, it is determined whether to recommend the target object to the user based on the user interest prediction result and the user conformity prediction result. For example, when the object popularity γ _f is set, the prediction result y _u,f can be determined according to the following formula:

in,

Predict results for user interests, and

Predict results for users following the crowd.

It should be noted that what is shown in Figure 6 is only an illustrative embodiment of the object recommendation process. In other embodiments, the object recommendation model may not have a graph neural network, so that the object feature embedding representation of the target object and the user interaction object does not perform graph enhancement processing based on the object knowledge graph.

Fig. 7 shows an example schematic diagram of an object recommendation process according to an embodiment of the present specification. In the example of Fig. 7, commodity recommendation is used as an example for explanation.

交互商品1的对象特征被提供给嵌入层得到交互商品1的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到交互商品1的经过图增强处理后的对象特征嵌入表征，交互商品2的对象特征被提供给嵌入层得到交互商品2的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到交互商品2的经过图增强处理后的对象特征嵌入表征，以及交互商品n的对象特征被提供给嵌入层得到交互商品n的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到交互商品n的经过图增强处理后的对象特征嵌入表征。目标商品的对象特征被提供给嵌入层得到目标商品的对象特征嵌入表征，并对所得到的对象特征嵌入表征经由图神经网络使用商品知识图谱进行图增强处理后得到目标商品的经过图增强处理后的对象特征嵌入表征x_f。目标对象的经过图增强处理后的对象特征嵌入表征被提供给对象推荐模型的表征层得到目标对象表征。As shown in Figure 7, the user features are provided to the embedding layer to obtain the user feature embedding representation

The object features of interactive product 1 are provided to the embedding layer to obtain the object feature embedding representation of interactive product 1, and the obtained object feature embedding representation is processed by graph neural network using product knowledge graph to obtain the object feature embedding representation of interactive product 1 after graph enhancement. The object features of interactive product 2 are provided to the embedding layer to obtain the object feature embedding representation of interactive product 2, and the obtained object feature embedding representation is processed by graph neural network using product knowledge graph to obtain the object feature embedding representation of interactive product 2 after graph enhancement. The object features of interactive product n are provided to the embedding layer to obtain the object feature embedding representation of interactive product n, and the obtained object feature embedding representation is processed by graph neural network using product knowledge graph to obtain the object feature embedding representation of interactive product n after graph enhancement. The object features of the target product are provided to the embedding layer to obtain the object feature embedding representation of the target product, and the obtained object feature embedding representation is processed by graph neural network using product knowledge graph to obtain the object feature embedding representation x _f of the target product after graph enhancement. The object feature embedding representation of the target object after graph enhancement is provided to the representation layer of the object recommendation model to obtain the target object representation.

将从交互商品1到交互商品n的对象特征嵌入表征中解耦出的与用户从众心理对应的对象特征嵌入表征分量级联， 得到用户从众嵌入表征

The object feature embedding representation components corresponding to user interests and user conformity are decoupled from the object feature embedding representation of interactive product 1 after graph enhancement. The object feature embedding representation components corresponding to user interests and user conformity are decoupled from the object feature embedding representation of interactive product 2 after graph enhancement. The object feature embedding representation components corresponding to user interests and user conformity are decoupled from the object feature embedding representation of interactive product n after graph enhancement. Then, the object feature embedding representation components corresponding to user interests decoupled from the object feature embedding representations of interactive product 1 to interactive product n are cascaded to obtain the user interest embedding representation

The object feature embedding representation components corresponding to the user's herd mentality that are decoupled from the object feature embedding representations from interactive product 1 to interactive product n are cascaded to obtain the user herd embedding representation

与用户特征嵌入表征

融合， 并将融合结果提供给前馈网络得到用户兴趣表征。将用户从众嵌入表征

与用户特征嵌入表征

融合，并将融合结果提供给前馈网络得到用户从众表征。然后，基于用户兴趣表征和目标对象表征得到用户兴趣预测结果

以及基于用户从众表征和目标对象表征得到用户从众预测结果

Then, user interests are embedded into the representation

Embedded representation with user features

Fusion, and provide the fusion result to the feedforward network to obtain the user interest representation.

Embedded representation with user features

Fusion, and provide the fusion result to the feedforward network to obtain the user conformity representation. Then, based on the user interest representation and the target object representation, the user interest prediction result is obtained.

And the user conformity prediction result is obtained based on the user conformity representation and the target object representation

和用户从众预测结果

后，基于用户兴趣预测结果

和用户从众预测结果

确定是否向用户推荐目标对象。After determining the user interest prediction results

and user herd prediction results

After that, based on the user interest prediction results

and user herd prediction results

Determine whether to recommend the target object to the user.

The object recommendation model training method and the object recommendation method according to the embodiments of the present specification are described above with reference to the accompanying drawings.

FIG8 shows an example block diagram of an object recommendation model training device 800 according to an embodiment of the present specification. As shown in FIG8 , the object recommendation model training device 800 includes a first representation generation unit 810, a graph enhancement processing unit 820, an embedding representation decoupling unit 830, a second representation generation unit 840, a target object prediction unit 850, a loss function determination unit 860, and a model parameter adjustment unit 870.

The first representation generation unit 810, the graph enhancement processing unit 820, the embedding representation decoupling unit 830, the second representation generation unit 840, the target object prediction unit 850, the loss function determination unit 860 and the model adjustment unit 870 perform operations in a loop until the training end condition is met.

Specifically, in each loop process, the first representation generation unit 810 is configured to provide the user feature, the object feature of the target object, the object feature of the user interaction object, and the object type of the user interaction object to the embedding layer of the object recommendation model, respectively, to obtain the user feature embedding representation of the user feature, the object feature embedding representation of the target object, the object feature embedding representation of the user interaction object, and the object type embedding representation of the user interaction object. Here, the user interaction object and the target object belong to the same type of object, and the object type has object risk information. The operation of the first representation generation unit 810 can refer to the operation described above with reference to 310 of Figure 3.

The graph enhancement processing unit 820 is configured to provide the object feature embedding representations of the target object and the user interaction object to the graph neural network of the object recommendation model to perform graph enhancement processing based on the object knowledge graph, and obtain the object feature embedding representations of the target object and the user interaction object after graph enhancement processing. The operation of the graph enhancement processing unit 820 can refer to the operation described above with reference to 320 of FIG. 3.

The embedding representation decoupling unit 830 is configured to decouple the user interest embedding representation, user conformity embedding representation and user risk preference embedding representation corresponding to the user interest, user conformity and user risk preference respectively from the object feature embedding representation of the user interaction object after the graph enhancement processing via the decoupling layer of the object recommendation model. The operation of the embedding representation decoupling unit 830 can refer to the operation described above with reference to 330 of FIG. 3 .

The second representation generation unit 840 is configured to fuse the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, and provide the obtained fusion results, the object feature embedding representation of the target object, and the object type embedding representation of the user interaction object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation, the target object representation, and the user risk preference representation. The operation of the second representation generation unit 840 can refer to the operations described above with reference to 340 and 350 of Figure 3.

The target object prediction unit 850 is configured to predict the user interest prediction result based on the user interest representation and the target object representation, and predict the user conformity prediction result based on the user conformity representation and the target object representation. The operation of the target object prediction unit 850 may refer to the operation described above with reference to 360 of FIG. 3 .

The loss function determination unit 860 is configured to determine the total loss function based on the user interest prediction result, the user conformity prediction result, the risk preference embedding representation, and the user risk preference representation in response to the training end condition not being met. The operation of the loss function determination unit 860 can refer to the operation described above with reference to 380 of FIG. 3.

The model adjustment unit 870 is configured to adjust the model parameters of the object recommendation model according to the total loss function. The operation of the model adjustment unit 870 may refer to the operation described above with reference to 390 of FIG. 3 .

FIG9 shows an example block diagram of a loss function determination unit 900 according to an embodiment of the present specification. As shown in FIG9 , the loss function determination unit 900 includes a user interest loss item determination module 910, a user conformity loss item determination module 920, a user risk preference loss item determination module 930, and a loss function determination module 940.

The user interest loss item determination module 910 is configured to determine the user interest loss item based on the user interest prediction result. The operation of the user interest loss item determination module 910 may refer to the operation described above with reference to 410 of FIG. 4 .

The user conformity loss term determination module 920 is configured to determine the user conformity loss term based on the user conformity prediction result. The operation of the user conformity loss term determination module 920 may refer to the operation described above with reference to 420 of FIG. 4 .

The user risk preference loss term determination module 930 is configured to determine the user risk preference loss term based on the risk preference embedding representation and the user risk preference representation. The operation of the user risk preference loss term determination module 930 may refer to the operation described above with reference to 430 of FIG. 4 .

The loss function determination module 940 is configured to determine a total loss function according to the user interest loss term, the user conformity loss term and the user risk preference loss term. The operation of the loss function determination module 940 may refer to the operation described above with reference to 440 of FIG. 4 .

In some embodiments, the embedding representation decoupling unit 830 can decouple the user interest embedding representation, user conformity embedding representation and user risk preference embedding representation corresponding to user interests, user conformity and user risk preferences respectively from the object feature embedding representation of the user interaction object based on an unsupervised mechanism via the decoupling layer of the object recommendation model.

In some embodiments, the decoupling layer may include a self-attention network. The embedding representation decoupling unit 830 may provide the object feature embedding representation of the user interaction object to the self-attention network to decouple the user interest embedding representation, user conformity embedding representation, and user risk preference embedding representation corresponding to the user interest, user conformity, and user risk preference, respectively, from the object feature embedding representation of the user interaction object.

In some embodiments, the user interaction object may include multiple user interaction objects. The embedding representation decoupling unit 830 is configured to decouple the feature embedding representation components corresponding to the user interest, user conformity and user risk preference from the object feature embedding representation of each user interaction object through the decoupling layer of the object recommendation model and cascade them to obtain the user interest embedding representation, the user conformity embedding representation and the user risk preference embedding representation.

It should be noted that Fig. 8 only shows an exemplary embodiment of the object recommendation model training device. In the example of Fig. 4, the object recommendation model is modeled using the user interest embedding representation, the user conformity embedding representation and the user risk preference embedding representation.

In some embodiments, only the user interest embedding representation and the user conformity embedding representation can be used to model the object recommendation model. In this case, when the model is trained, the first representation generation unit provides the user features, the object features of the target object, and the object features of the user interaction object to the embedding layer of the object recommendation model, respectively, to obtain the user feature embedding representation of the user features, the object feature embedding representation of the target object, and the object feature embedding representation of the user interaction object. The embedding representation decoupling unit decouples the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology from the object feature embedding representation of the user interaction object after the graph enhancement processing through the decoupling layer of the object recommendation model. The second representation generation unit fuses the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, respectively, and provides the obtained fusion results and the object feature embedding representation of the target object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation, and the target object representation. Subsequently, the loss function determination unit determines the total loss function based on the user interest prediction result and the user conformity prediction result. Specifically, the loss function determination unit determines the user interest loss term based on the user interest prediction result, and determines the user conformity loss term based on the user conformity prediction result. Then, the total loss function is determined based on the user interest loss term and the user conformity loss term.

In some embodiments, the object recommendation model may not include a graph neural network. Accordingly, the object recommendation model training device may not include a graph enhancement processing unit, so as not to perform graph enhancement processing based on the object knowledge graph on the object feature embedding representation of the target object and the user interaction object.

Fig. 10 shows an example block diagram of an object recommendation apparatus 1000 according to an embodiment of the present specification. As shown in Fig. 10, the object recommendation apparatus 1000 includes a first representation generation unit 1010, a graph enhancement processing unit 1020, an embedding representation decoupling unit 1030, a second representation generation unit 1040 and a recommendation decision unit 1050.

The first representation generation unit 1010 is configured to provide the user features, the object features of the target object, and the object features of the user interaction object to the embedding layer of the object recommendation model to obtain the user feature embedding representation of the user features, and the object feature embedding representation of the target object and the user interaction object. The operation of the first representation generation unit 1010 may refer to the operation described above with reference to 610 of FIG. 6 .

The graph enhancement processing unit 1020 is configured to provide the object feature embedding representation of the target object and the object feature embedding representation of the user interaction object to the graph neural network of the object recommendation model to perform graph enhancement processing based on the object knowledge graph, and obtain the object feature embedding representation of the target object and the user interaction object after graph enhancement processing. The operation of the graph enhancement processing unit 1020 can refer to the operation described above with reference to 620 of FIG. 6.

The embedding representation decoupling unit 1030 is configured to decouple the user interest embedding representation and the user conformity embedding representation corresponding to the user interest and the user conformity psychology respectively from the object feature embedding representation of the user interaction object after the graph enhancement processing via the decoupling layer of the object recommendation model. In some embodiments, the user interaction object may include multiple user interaction objects. The embedding representation decoupling unit 1030 is configured to decouple the feature embedding representation components corresponding to the user interest, the user conformity psychology and the user risk preference respectively from the object feature embedding representation of each user interaction object after the graph enhancement processing via the decoupling layer of the object recommendation model and cascade them to obtain the user interest embedding representation, the user conformity embedding representation and the user risk preference embedding representation. The operation of the embedding representation decoupling unit 1030 can refer to the operation described above with reference to 630 of Figure 6.

The second representation generation unit 1040 is configured to fuse the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation, and provide the obtained fusion results and the object feature embedding representation of the target object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation and the target object representation. The operation of the second representation generation unit 1040 can refer to the operations described above with reference to 640 and 650 of FIG. 6.

The recommendation decision unit 1050 is configured to determine whether to recommend the target object to the user based on the user interest representation, the user conformity representation, and the target object representation. For example, the recommendation decision unit 1050 can predict the user interest prediction result based on the user interest representation and the target object representation, and predict the user conformity prediction result based on the user conformity representation and the target object representation, and determine whether to recommend the target object to the user based on the user interest prediction result and the user conformity prediction result. The operation of the recommendation decision unit 1050 can refer to the operations described above with reference to 660 and 670 of Figure 6.

As described above with reference to Figures 1 to 10, the object recommendation model training method, object recommendation model training device, object recommendation method and object recommendation device according to the embodiments of this specification are described. The above object recommendation model training device and object recommendation device can be implemented by hardware, or by software or a combination of hardware and software.

FIG11 shows an example schematic diagram of an object recommendation model training device 1100 implemented based on a computer system according to an embodiment of the present specification. As shown in FIG11 , the object recommendation model training device 1100 may include at least one processor 1110, a memory (e.g., a non-volatile memory) 1120, a memory 1130, and a communication interface 1140, and the at least one processor 1110, the memory 1120, the memory 1130, and the communication interface 1140 are connected together via a bus 1160. At least one processor 1110 executes at least one computer-readable instruction stored or encoded in the memory (i.e., the above-mentioned element implemented in software form).

In one embodiment, a computer executable instruction is stored in a memory, which, when executed, causes at least one processor 1110 to: loop through the following model training process until a training end condition is met: provide user features, object features of user interaction objects, and object features of target objects to an embedding layer of an object recommendation model, respectively, to obtain a user feature embedding representation of the user features, an object feature embedding representation of the user interaction objects, and an object feature embedding representation of the target object; decouple, via a decoupling layer of the object recommendation model, a user interest embedding representation and a user conformity embedding representation corresponding to user interests and user conformity, respectively, from the object feature embedding representation of the user interaction objects Representation; fusing the user feature embedding representation with the user interest embedding representation and the user conformity embedding representation respectively, and providing the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model to obtain the user interest representation, the user conformity representation and the target object representation; predicting the user interest prediction result based on the user interest representation and the target object representation, and predicting the user conformity prediction result based on the user conformity representation and the target object representation; and in response to not meeting the training end condition, determining the total loss function based on the user interest prediction result and the user conformity prediction result, and adjusting the model parameters of the object recommendation model according to the total loss function.

It should be understood that the computer executable instructions stored in the memory, when executed, enable at least one processor 1110 to perform various operations and functions described above in conjunction with Figures 1-5 and Figures 8-9 in various embodiments of this specification.

FIG12 shows an example schematic diagram of an object recommendation device 1200 implemented based on a computer system according to an embodiment of the present specification. As shown in FIG12 , the object recommendation device 1200 may include at least one processor 1210, a memory (e.g., a non-volatile memory) 1220, a memory 1230, and a communication interface 1240, and the at least one processor 1210, the memory 1220, the memory 1230, and the communication interface 1240 are connected together via a bus 1260. At least one processor 1210 executes at least one computer-readable instruction stored or encoded in the memory (i.e., the above-mentioned element implemented in the form of software).

In one embodiment, computer executable instructions are stored in a memory, which when executed cause at least one processor 1210 to: provide user features, object features of user interaction objects, and object features of target objects to an embedding layer of an object recommendation model, respectively, to obtain user feature embedding representations of user features, object feature embedding representations of user interaction objects, and object feature embedding representations of target objects; decouple user interest embedding representations and user conformity embedding representations, respectively corresponding to user interests and user herd mentality, from the object feature embedding representations of user interaction objects via a decoupling layer of the object recommendation model; fuse the user feature embedding representation with the user interest embedding representation and the user herd embedding representation, respectively, and provide the obtained fusion result and the object feature embedding representation of the target object to the representation layer of the object recommendation model, respectively, to obtain user interest representation, user herd representation, and target object representation; and determine whether to recommend the target object to the user based on the user interest representation, the user herd representation, and the target object representation.

It should be understood that the computer executable instructions stored in the memory, when executed, enable at least one processor 1210 to perform the various operations and functions described above in conjunction with Figures 6-7 and Figure 10 in the various embodiments of this specification.

According to one embodiment, a program product such as a machine-readable medium (e.g., a non-transitory machine-readable medium) is provided. The machine-readable medium may have instructions (i.e., the above-mentioned elements implemented in software form), which, when executed by a machine, causes the machine to perform the various operations and functions described above in conjunction with Figures 1-10 in the various embodiments of this specification. Specifically, a system or device equipped with a readable storage medium may be provided, on which a software program code that implements the functions of any of the above-mentioned embodiments is stored, and a computer or processor of the system or device is caused to read and execute instructions stored in the readable storage medium.

In this case, the program code itself read from the machine-readable medium can realize the function of any one of the above-mentioned embodiments, and thus the machine-readable code and the machine-readable storage medium storing the machine-readable code constitute a part of the present invention.

Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or a cloud via a communication network.

According to one embodiment, a computer program product is provided, which includes a computer program. When the computer program is executed by a processor, the processor performs the various operations and functions described above in conjunction with Figures 1 to 10 in the various embodiments of this specification.

Those skilled in the art should understand that the various embodiments disclosed above can be modified and altered in various ways without departing from the essence of the invention. Therefore, the protection scope of the present invention should be defined by the appended claims.

It should be noted that not all steps and units in the above-mentioned processes and system structure diagrams are necessary, and some steps or units can be ignored according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or some components in multiple independent devices may be implemented together.

In the above embodiments, the hardware unit or module can be realized by mechanical or electrical means. For example, a hardware unit, module or processor can include permanent dedicated circuit or logic (such as special processor, FPGA or ASIC) to complete the corresponding operation. The hardware unit or processor can also include programmable logic or circuit (such as general-purpose processor or other programmable processor), which can be temporarily set by software to complete the corresponding operation. Specific implementation (mechanical method or dedicated permanent circuit or temporary circuit) can be determined based on cost and time consideration.

The specific embodiments described above in conjunction with the accompanying drawings describe exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of protection of the claims. The term "exemplary" used throughout this specification means "used as an example, instance or illustration" and does not mean "preferred" or "having advantages" over other embodiments. For the purpose of providing an understanding of the described technology, the specific embodiments include specific details. However, these technologies can be implemented without these specific details. In some instances, in order to avoid making the concepts of the described embodiments difficult to understand, well-known structures and devices are shown in block diagram form.

The above description of the disclosure is provided to enable any person of ordinary skill in the art to implement or use the disclosure. Various modifications to the disclosure will be apparent to those of ordinary skill in the art, and the general principles defined herein may be applied to other variations without departing from the scope of protection of the disclosure. Therefore, the disclosure is not limited to the examples and designs described herein, but is consistent with the widest range of principles and novel features disclosed herein.

Claims (21)

1. A method for training an object recommendation model for determining whether to recommend a target object to a user, the method comprising:

the following model training process is circularly executed until the training ending condition is met:

providing the user characteristics, the object characteristics of the user interaction object and the object characteristics of the target object to an embedding layer of the object recommendation model respectively to obtain user characteristic embedding characterization of the user characteristics, object characteristic embedding characterization of the user interaction object and object characteristic embedding characterization of the target object;

decoupling the user interest embedded representation and the user popular embedded representation which respectively correspond to the user interests and the user popular psychology from the object feature embedded representation of the user interaction object through a decoupling layer of the object recommendation model;

The user characteristic embedded characterization is fused with the user interest embedded characterization and the user popular embedded characterization respectively, and the obtained fusion result and the object characteristic embedded characterization of the target object are provided for a characterization layer of the object recommendation model respectively to obtain user interest characterization, user popular characterization and target object characterization;

predicting a user interest prediction result based on the user interest representation and the target object representation, and predicting a user preference prediction result based on the user preference representation and the target object representation; and

and determining a total loss function based on the user interest prediction result and the user follow-up prediction result in response to the training ending condition not being met, and adjusting model parameters of the object recommendation model according to the total loss function.

2. The method of claim 1, further comprising:

and embedding object features of the target object and the user interaction object into a graph neural network, wherein the graph neural network is characterized in that the object features are respectively provided for the object recommendation model, so that graph enhancement processing is performed on the basis of an object knowledge graph, and the object features of the target object and the user interaction object after the graph enhancement processing are obtained.

3. The method of claim 1, wherein the user interaction object comprises at least two user interaction objects,

decoupling, via a decoupling layer of the object recommendation model, the user interest embedding representation and the user popular embedding representation corresponding to the user interests and the user popular psychology, respectively, from the object feature embedding representation of the user interaction object comprises:

and decoupling object feature embedded characterization components corresponding to the user interests and the user psychology respectively from object feature embedded characterization of each user interaction object through a decoupling layer of the object recommendation model, and cascading to obtain user interest embedded characterization and user psychology embedded characterization.

4. The method of claim 1, wherein decoupling, via a decoupling layer of the object recommendation model, the user interest embedded representation and the user popular embedded representation corresponding to the user interests and the user popular psychology, respectively, from the object feature embedded representation of the user interaction object comprises:

decoupling the user interest embedded representation and the user popular embedded representation corresponding to the user interests and the user popular psychology respectively from the object feature embedded representation of the user interaction object based on an unsupervised mechanism via a decoupling layer of the object recommendation model.

5. The method of claim 4, wherein the decoupling layer comprises a self-attention network, and decoupling, via the decoupling layer of the object recommendation model, the user interest embedded representation and the user popular embedded representation, respectively corresponding to the user interests and the user popular psychology, from the object feature embedded representation of the user interaction object comprises:

and providing the object feature embedded representation of the user interaction object to the self-attention network so as to decouple the user interest embedded representation and the user popular embedded representation which respectively correspond to the user interests and the user popular psychology from the object feature embedded representation of the user interaction object.

6. The method of any of claims 1 to 5, wherein determining a total loss function based on the user interest prediction result and the user follow-up prediction result comprises:

determining a user interest loss term based on the user interest prediction result;

determining a user audience loss term based on the user audience prediction results; and

and determining the total loss function according to the user interest loss term and the user audience loss term.

7. The method of claim 6, wherein the target object has an object popularity determined based on a number of interactions in an object interaction history behavior, the user interest loss term is determined based on the user interest prediction result and the object popularity of the target object, and the user audience loss term is determined based on the user audience prediction result and the object popularity of the target object.

8. The method according to claim 1 to 5, wherein,

providing the user characteristics, the object characteristics of the user interaction object and the object characteristics of the target object to an embedding layer of the object recommendation model respectively, and obtaining the user characteristic embedding characterization of the user characteristics, the object characteristic embedding characterization of the user interaction object and the object characteristic embedding characterization of the target object comprises:

providing the user characteristics, the object characteristics of the user interaction object, the object characteristics of the target object and the object types of the user interaction object to an embedding layer of the object recommendation model respectively to obtain the user characteristic embedding characterization of the user characteristics, the object characteristic embedding characterization of the user interaction object, the object characteristic embedding characterization of the target object and the object type embedding characterization of the user interaction object,

decoupling, via a decoupling layer of the object recommendation model, the user interest embedding representation and the user popular embedding representation corresponding to the user interests and the user popular psychology, respectively, from the object feature embedding representation of the user interaction object comprises:

decoupling, via a decoupling layer of the object recommendation model, from object feature embedding characterizations of the user interaction object, user interest embedding characterizations, user popular embedding characterizations and user risk preference embedding characterizations corresponding to user interests, user popular psychology and user risk preferences, respectively,

The method for obtaining the user interest characterization, the user secondary characterization and the target object characterization by respectively providing the obtained fusion result and the object feature embedded characterization of the target object to the characterization layer of the object recommendation model comprises the following steps:

providing the obtained fusion result, the object feature embedded representation of the target object and the object type embedded representation of the user interaction object to a representation layer of the object recommendation model to obtain a user interest representation, a user secondary representation, a target object representation and a user risk preference representation,

determining a total loss function based on the user interest prediction result and the user-friendly prediction result includes:

a total loss function is determined based on the user interest prediction result, the user preference embedding characterization, the user preference characterization, and the user preference characterization.

9. The method of claim 8, wherein determining a total loss function based on the user interest prediction result, the user follow-up prediction result, the risk preference embedded characterization, and the user risk preference characterization comprises:

determining a user interest loss term based on the user interest prediction result;

determining a user audience loss term based on the user audience prediction results;

Determining a user risk preference loss term based on the risk preference embedded representation and the user risk preference representation; and

and determining the total loss function according to the user interest loss item, the user audience loss item and the user risk preference loss item.

10. The method of claim 9, wherein the user risk preference penalty term has a weighted value defining a degree of contribution of the user risk preference penalty term to the overall penalty function,

determining the total loss function from the user interest loss term, the user audience loss term, and the user risk preference loss term comprises:

and determining the total loss function according to the user interest loss term, the user audience loss term and the weighted user risk preference loss term.

11. The method of claim 2, wherein the object knowledge graph is created based on a relationship between a recommendation type object and an association type object.

12. A method for determining a recommended object based on an object recommendation model, comprising:

providing the user characteristics, the object characteristics of the user interaction object and the object characteristics of the target object to an embedding layer of the object recommendation model respectively to obtain user characteristic embedding characterization of the user characteristics, object characteristic embedding characterization of the user interaction object and object characteristic embedding characterization of the target object;

Decoupling the user interest embedded representation and the user popular embedded representation which respectively correspond to the user interests and the user popular psychology from the object feature embedded representation of the user interaction object through a decoupling layer of the object recommendation model;

the user characteristic embedded characterization is fused with the user interest embedded characterization and the user popular embedded characterization respectively, and the obtained fusion result and the object characteristic embedded characterization of the target object are provided for a characterization layer of the object recommendation model respectively to obtain user interest characterization, user popular characterization and target object characterization; and

based on the user interest characterization, the user audience characterization, and the target object characterization, it is determined whether to recommend the target object to a user.

13. The method of claim 12, wherein determining whether to recommend the target object to a user based on the user interest characterization, the user audience characterization, and the target object characterization comprises:

predicting a user interest prediction result and a user preference prediction result based on the user interest characterization and the target object characterization and based on the user preference characterization and the target object characterization, respectively; and

And determining whether to recommend the target object to the user according to the user interest prediction result and the user popular prediction result.

14. The method of claim 12, further comprising:

and embedding object features of the target object and the user interaction object into a graph neural network, wherein the graph neural network is characterized in that the object features are respectively provided for the object recommendation model, so that graph enhancement processing is performed on the basis of an object knowledge graph, and the object features of the target object and the user interaction object after the graph enhancement processing are obtained.

15. The method of claim 12, wherein the object recommendation model is trained in accordance with the method of any one of claims 1 to 11.

16. An apparatus for training an object recommendation model for determining whether to recommend a target object to a user, the apparatus comprising:

the first characterization generating unit is used for respectively providing the user characteristics, the object characteristics of the user interaction object and the object characteristics of the target object to an embedding layer of the object recommendation model to obtain user characteristic embedding characterization of the user characteristics, object characteristic embedding characterization of the user interaction object and object characteristic embedding characterization of the target object;

The embedded representation decoupling unit is used for decoupling the user interest embedded representation and the user popular embedded representation which respectively correspond to the user interests and the user popular psychology from the object feature embedded representation of the user interaction object through a decoupling layer of the object recommendation model;

the second characterization generating unit is used for respectively fusing the user characteristic embedded characterization with the user interest embedded characterization and the user popular embedded characterization, and respectively providing the obtained fusion result and the object characteristic embedded characterization of the target object for a characterization layer of the object recommendation model to obtain the user interest characterization, the user popular characterization and the target object characterization;

a target object prediction unit that predicts a user interest prediction result based on the user interest characterization and the target object characterization, and predicts a user-by-user prediction result based on the user-by-user characterization and the target object characterization;

a loss function determining unit that determines a total loss function based on the user interest prediction result and the user-popular prediction result in response to the training end condition not being satisfied; and

a model adjustment unit for adjusting model parameters of the object recommendation model according to the total loss function,

The first characterization generating unit, the graph enhancement processing unit, the embedded characterization decoupling unit, the second characterization generating unit, the third characterization generating unit, the target object predicting unit, the loss function determining unit and the model adjusting unit circularly execute operations until the training ending condition is met.

17. An apparatus for determining a recommended object based on an object recommendation model, comprising:

the first characterization generating unit is used for respectively providing the user characteristics, the object characteristics of the user interaction object and the object characteristics of the target object to an embedding layer of the object recommendation model to obtain user characteristic embedding characterization of the user characteristics, object characteristic embedding characterization of the user interaction object and object characteristic embedding characterization of the target object;

the embedded representation decoupling unit is used for decoupling the user interest embedded representation and the user popular embedded representation which respectively correspond to the user interests and the user popular psychology from the object feature embedded representation of the user interaction object through a decoupling layer of the object recommendation model;

the second characterization generating unit is used for respectively fusing the user characteristic embedded characterization with the user interest embedded characterization and the user popular embedded characterization, and respectively providing the obtained fusion result and the object characteristic embedded characterization of the target object for a characterization layer of the object recommendation model to obtain the user interest characterization, the user popular characterization and the target object characterization; and

And a recommendation decision unit for determining whether to recommend the target object to the user based on the user interest characterization, the user audience characterization and the target object characterization.

18. The apparatus of claim 17, further comprising:

and the graph enhancement processing unit is used for embedding the object characteristics of the target object and the user interaction object into the graph neural network, the object characteristics of which are respectively provided for the object recommendation model, so as to perform graph enhancement processing based on the object knowledge graph, and obtain the object characteristic embedding characteristics of the target object and the user interaction object after the graph enhancement processing.

19. An object recommendation system, comprising:

the interactive object acquisition device acquires a user interactive object;

the apparatus for training an object recommendation model of claim 16; and

the apparatus for determining recommended objects based on an object recommendation model according to claim 17 or 18.

20. An apparatus for training an object recommendation model, comprising:

at least one of the processors is configured to perform,

a memory coupled to the at least one processor, and

computer program stored in the memory, the at least one processor executing the computer program to implement the method for training an object recommendation model according to any of claims 1 to 11.

21. An apparatus for determining a recommended object based on an object recommendation model, comprising:

at least one of the processors is configured to perform,

a memory coupled to the at least one processor, and

computer program stored in the memory, the at least one processor executing the computer program to implement the method for determining recommended objects based on an object recommendation model as claimed in any one of claims 12 to 15.

Images