CN114491291A - User preference prediction method, terminal and storage medium - Google Patents

Abstract

The invention provides a user preference prediction method, terminal and storage medium. The method includes: constructing an explicit scoring matrix and an implicit scoring matrix according to the historical interaction data of m users and n items, and the explicit scoring matrix is used for Represents the explicit ratings of m users on n items, and the implicit rating matrix is used to represent the implicit ratings of m users on n items. Both the explicit rating matrix and the implicit rating matrix are matrices with m rows and n columns; Input the explicit scoring matrix and the implicit scoring matrix into the pre-set matrix decomposition and deep neural network joint prediction model, and obtain the prediction result. The item's preference value. The present invention can improve the prediction accuracy of user preference.

Description

User preference prediction method, terminal and storage medium

technical field

The present invention relates to the technical field of deep learning, and in particular, to a user preference prediction method, a terminal and a storage medium.

Background technique

In recent years, the rapid development of social network media industry and e-commerce industry has greatly increased the number of Internet users participating. Faced with the massive multi-source heterogeneous data generated when users browse the Internet, how to process it accurately and effectively to improve the recommendation accuracy and user satisfaction is still a research hotspot of personalized recommendation services.

However, existing prediction methods have low prediction accuracy and low user satisfaction.

SUMMARY OF THE INVENTION

In view of this, the present invention provides a user preference prediction method, terminal and storage medium, which can solve the problem of low accuracy of user preference prediction.

In a first aspect, an embodiment of the present invention provides a user preference prediction method, including:

According to the historical interaction data of m users and n items, an explicit scoring matrix and an implicit scoring matrix are constructed. The explicit scoring matrix is used to represent the explicit scoring of the n items by the m users. The implicit scoring matrix is used to represent the implicit scoring of the n items by the m users, and both the explicit scoring matrix and the implicit scoring matrix are matrices with m rows and n columns;

Inputting the explicit scoring matrix and the implicit scoring matrix into a preset matrix decomposition and a deep neural network joint prediction model to obtain a prediction result, for each of the m users, the prediction result includes the The value of the user's preference for each of the n items.

In a possible implementation, inputting the explicit scoring matrix and the implicit scoring matrix into a preset matrix decomposition and deep neural network joint prediction model, and obtaining the prediction result includes:

According to the explicit scoring matrix and the implicit scoring matrix, determine the potential feature vector of the user and the potential feature vector of the item, respectively;

According to the latent feature vector of the user and the latent feature vector of the item, determine the shallow linear preference feature and the deep nonlinear preference feature of the m users for the n items;

According to the shallow linear preference feature and the deep nonlinear preference feature of the m users for the n items, the prediction result is obtained through a preset activation function.

In a possible implementation manner, the determining of the potential feature vector of the user and the potential feature vector of the item respectively according to the explicit scoring matrix and the implicit scoring matrix includes:

According to the displayed rating matrix and the implicit rating matrix, through normal random initialization, the explicit latent eigenvector and the implicit latent eigenvector of the user, and the explicit latent eigenvector and the implicit latent eigenvector of the item are obtained;

adding the user's explicit latent feature vector and the implicit latent feature vector to obtain the user's latent feature vector;

The item's explicit latent feature vector and the implicit latent feature vector are added to obtain the item's latent feature vector.

In a possible implementation, according to the potential feature vector of the user and the potential feature vector of the item, determine the shallow linear preference feature and the deep nonlinear preference of the m users for the n items Preference features include:

The dot product calculation is performed on the potential feature vector of the user and the potential feature vector of the item, as the input of a preset matrix factorization model, and the matrix factorization model outputs the shallow results of the m users on the n items. Layer linear preference feature;

The potential feature vector of the user and the potential feature vector of the item are used as the input of the preset deep neural network model, and the hidden layer of the multi-layer perceptron of the deep neural network model is used to obtain the information about the m users. The deep nonlinear preference feature of the n items, wherein the matrix decomposition and deep neural network joint prediction model includes the matrix decomposition model and the deep neural network model.

In a possible implementation, the loss function of the matrix factorization and the deep neural network joint prediction model is

L= _{ηL I} +(1-η)LE

Wherein, L is the loss function of the matrix decomposition and the deep neural network joint prediction model, L _E is the loss function of the explicit score, L _I is the loss function of the implicit score, n is the preset coefficient, n is greater than 0 and less than 1.

In a possible implementation manner, the loss function of the displayed score is

为所述矩阵分解与深度神经网络联合预测模型的预测值；Wherein, L _E is the loss function of the explicit scoring, UE is the number of users in the explicit scoring matrix, _IE is the number of items in the explicit scoring matrix, and r _ui is the user u to item _i The actual value of the explicit score, Max(R) is the maximum value of the explicit score,

is the predicted value of the combined prediction model of the matrix decomposition and the deep neural network;

The loss function of the implicit score is

为矩阵分解与深度神经网络联合预测模型的预测值。Among them, L _I is the loss function of implicit scoring, U _I is the number of users in the implicit scoring matrix, I _I is the number of items in the implicit scoring matrix, ir _ui is the implicit score of user u on item i rating value,

is the prediction value of the combined prediction model of matrix factorization and deep neural network.

In a possible implementation manner, before constructing the explicit scoring matrix and the implicit scoring matrix according to the historical interaction data of m users and n items, the method further includes:

Remove the abnormal data in the initial data set to obtain the target data set, wherein the initial data set includes historical interaction data of multiple users and multiple items, for each user, if the user and multiple items in the initial data set The number of historical interactions is less than or equal to the first preset number of times, then the user's historical interaction data is abnormal data, and, for each item, if the number of historical interactions between multiple users in the initial data set and the item is less than or equal to the first Two preset times, the historical interaction data of the item is abnormal data, the first preset times and the second preset times are the same or different, and the target data set includes the m users and n items historical interaction data.

In a possible implementation manner, the constructing the explicit scoring matrix and the implicit scoring matrix according to the historical interaction data between m users and n items includes:

If user u scores item i, then normalize user u's score for item i to determine the normalized score value; if user u does not score item i, then use user u's score for item i Set to 0, construct the explicit score matrix according to the score of each user for each item, and for any normalized score value, the score value is greater than or equal to 0 and less than or equal to 1;

If user u interacts with item i, set user u's implicit score to item i to 1; if user u has no interaction with item i, set user u's implicit score to item i to 0, according to The implicit rating of each user for each item, constructing the implicit rating matrix.

In a second aspect, an embodiment of the present invention provides a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, which is implemented when the processor executes the computer program The steps of the method described above in the first aspect or any possible implementation manner of the first aspect.

In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, implements the first aspect or any of the first aspect above. A possible implementation of the steps of the described method.

The beneficial effects that the embodiment of the present invention has compared with the prior art are:

In the embodiment of the present invention, on the one hand, an explicit scoring matrix and an implicit scoring matrix are constructed in combination with the explicit scoring and implicit scoring of the item by the user, and the explicit scoring matrix and the implicit scoring matrix are used as the input of the prediction model , make full use of the respective features of explicit scoring and implicit scoring and the complementarity of the combination of the two to make it suitable for scoring scenarios. Combining the advantages of the two models further improves the accuracy of user preference prediction. Therefore, the method provided by the implementation of the present invention can improve the accuracy of the recommendation system in predicting the user's preference for an item, and display a personalized item recommendation list for the user. In order to provide users with more accurate personalized services, improve user satisfaction.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only for the present invention. In some embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is the realization flow chart of a kind of user preference prediction method provided by the embodiment of the present invention;

FIG. 2 is an implementation flowchart of another user preference prediction method provided by an embodiment of the present invention;

3 is a schematic structural diagram of an apparatus for predicting user preference provided by an embodiment of the present invention;

FIG. 4 is a schematic diagram of a terminal provided by an embodiment of the present invention.

Detailed ways

In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are set forth in order to provide a thorough understanding of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

In order to make the objectives, technical solutions and advantages of the present invention clearer, the following descriptions will be given through specific embodiments in conjunction with the accompanying drawings.

Existing personalized recommendation systems generally build recommendation models through historical interaction data generated when users browse websites. There are two main types of data: explicit feedback data and implicit feedback data. Explicit feedback data is generally data such as ratings, which can accurately reflect user preferences; implicit feedback is mostly user browsing lines such as clicks and favorites, which can be converted into binary data. Although it cannot accurately distinguish user preferences, it can reflect user preferences. underlying interest preferences. However, in the prior art, a recommendation model is usually constructed from a single user implicit feedback data, or a recommendation model is constructed from a single user explicit feedback data. , resulting in a low prediction accuracy of user preference. To solve this problem, an embodiment of the present invention provides a user preference prediction method. Referring to FIG. 1 , it shows the implementation of the user preference prediction method provided by the embodiment of the present invention. Flowchart, detailed as follows:

In step 101, an explicit scoring matrix and an implicit scoring matrix are constructed according to the historical interaction data between m users and n items. The explicit scoring matrix is used to represent the explicit scoring of m users on n items, and the implicit scoring matrix The rating matrix is used to represent the implicit ratings of m users on n items. Both the explicit rating matrix and the implicit rating matrix are matrices with m rows and n columns.

Among them, m is a positive integer greater than or equal to 2, and n is a positive integer greater than or equal to 2.

In a possible implementation, the data set is too sparse due to the users and items with too few interactions in the dataset, resulting in low prediction accuracy of the model. To solve this problem, according to the historical interactions between m users and n items data, before constructing the explicit scoring matrix and the implicit scoring matrix, the method provided by the embodiment of the present invention further includes:

Remove the abnormal data in the initial data set to obtain the target data set, wherein the initial data set includes the historical interaction data of multiple users and multiple items. is less than or equal to the first preset number of times, the user's historical interaction data is abnormal data, and, for each item, if the number of historical interactions between multiple users in the initial data set and the item is less than or equal to the second preset number of times, then The historical interaction data of the item is abnormal data, the first preset number of times is the same or different from the second preset number of times, and the target data set includes historical interaction data of m users and n items.

For example, the first preset number of times is 2, and the second preset number of times is 5. In the initial data set, if there is user A, the number of interactions between user A and multiple items in the initial data set is only 1, that is, the user If you have only visited one of the items, or have scored on one of the items, the number of historical interactions of user A is too low, and the historical interaction data of user A is abnormal data, and the historical interaction data of user A is deleted from the initial data set. For another example, in the initial data set, if there is item X, the number of interactions between item X and multiple users in the initial data set is only 3 times, that is, the item value has been accessed and/or evaluated three times, then the historical interaction times of item X If it is too low, the historical interaction data of item X is abnormal data.

In a possible implementation, according to the historical interaction data of m users and n items, constructing an explicit scoring matrix and an implicit scoring matrix includes:

If user u scores item i, then normalize user u's score for item i to determine the normalized score value; if user u does not score item i, then use user u's score for item i Set to 0, construct an explicit score matrix according to the score of each user for each item, for any normalized score value, the score value is greater than or equal to 0 and less than or equal to 1;

If user u interacts with item i, set user u's implicit score to item i to 1; if user u has no interaction with item i, set user u's implicit score to item i to 0, according to The implicit rating of each user for each item, constructing an implicit rating matrix.

For example, in the target dataset, the historical interaction data of m users and n items is included. The user set is U={u ₁ u ₂ ,..., _um }, and the item set is V={v ₁ ,v ₂ ,...,v _n }. According to the characteristics of explicit and implicit feedback in the historical interaction data between users and items, construct an explicit rating matrix R=[r _ui ] _m×n for users and items and an implicit rating matrix IR=[ir _ui ] _m ×n for users and items _n .

In step 102, the explicit scoring matrix and the implicit scoring matrix are input into the preset matrix decomposition and deep neural network joint prediction model to obtain a prediction result. The value of the degree of preference for each of the items.

Matrix Factorization (MF) is a model-based recommendation method, which has the characteristics of strong scalability and low complexity. It uses latent factor vectors to represent users and items, maps users and items to a joint low-dimensional latent space, and formulates recommendation as a user preference prediction problem for items based on the linear inner product of the corresponding user and item latent factor vectors. However, the matrix factorization recommendation method only obtains the shallow linear latent factor feature vectors of users and items, and cannot well model the complex nonlinear deep feature representation between users and items.

With the rapid development and application of deep learning technology, the prior art proposes a recommendation method based on deep learning to model complex nonlinear interactions between users and items, so as to predict user preferences. But at present, most deep learning recommendation models model the attribute data of interacting users and the item itself, as well as auxiliary information such as social network and geographic location, and only use one of explicit feedback data or implicit feedback data for feedback. Modeling of user-item interactions to predict user preferences. Failure to take full advantage of the user-item interaction rating data and the complementarity of explicit rating and implicit feedback to improve the performance of recommender systems. The two types of heterogeneous data are complementary, and the combined recommendation model can effectively improve the recommendation effect.

In the embodiment of the present invention, on the one hand, an explicit scoring matrix and an implicit scoring matrix are constructed in combination with the explicit scoring and implicit scoring of the item by the user, and the explicit scoring matrix and the implicit scoring matrix are used as the input of the prediction model , make full use of the respective features of explicit scoring and implicit scoring and the complementarity of the combination of the two to make it suitable for scoring scenarios. Combining the advantages of the two models further improves the accuracy of user preference prediction. Therefore, the method provided by the implementation of the present invention can improve the accuracy of the recommendation system in predicting the user's preference for an item, and display a personalized item recommendation list for the user. In order to provide users with more accurate personalized services, improve user satisfaction.

FIG. 2 is an implementation flowchart of another user preference prediction method provided by an embodiment of the present invention, which is described in detail as follows:

In step 201, an explicit scoring matrix and an implicit scoring matrix are constructed according to the historical interaction data between m users and n items. The explicit scoring matrix is used to represent the explicit scoring of m users on n items, and the implicit scoring matrix The rating matrix is used to represent the implicit ratings of m users on n items. Both the explicit rating matrix and the implicit rating matrix are matrices with m rows and n columns.

For a specific implementation manner of step 201, reference may be made to step 101 in the embodiment corresponding to FIG. 1 , which is not repeated in this embodiment of the present invention.

In step 202, the potential feature vector of the user and the potential feature vector of the item are determined according to the explicit scoring matrix and the implicit scoring matrix, respectively.

In the embodiment of the present invention, the matrix decomposition and deep neural network joint prediction model includes an input layer, an Embedding layer, a mixed model layer, and a prediction layer. In the embodiment of the present invention, the matrix decomposition and the deep neural network joint prediction model and the joint prediction model are used to represent the same concept.

In the input layer of the joint prediction model, the explicit scoring matrix and the implicit scoring matrix are input into the matrix decomposition and deep neural network joint prediction model. In order to solve the problem that the explicit scoring matrix and the implicit scoring matrix are too sparse, in the embodiment of the present invention, the Embedding layer of the joint prediction model uses the dimension reduction idea for the explicit scoring matrix and the implicit scoring matrix to obtain the potential feature vector of the user. and latent feature vectors of items.

In a possible implementation manner, according to the explicit scoring matrix and the implicit scoring matrix, respectively determining the potential feature vector of the user and the potential feature vector of the item includes:

According to the explicit rating matrix and the implicit rating matrix, through normal random initialization, the explicit latent eigenvectors and implicit latent eigenvectors of users, as well as the explicit latent eigenvectors and implicit latent eigenvectors of items, are obtained; The latent feature vector and the implicit latent feature vector are added to obtain the user's latent feature vector; the item's explicit latent feature vector and the implicit latent feature vector are added to obtain the item's latent feature vector.

和隐式潜在特征向量

以及项目的显式潜在特征向量

和隐式潜在特征向量

将显式潜在特征向量和隐式潜在特征向量相加互为补充，得到用户的潜在特征向量

和项目的潜在特征向量

Using normal random initialization to obtain explicit latent feature vectors of users

and the implicit latent feature vector

and the explicit latent feature vector of the item

and the implicit latent feature vector

Add the explicit latent feature vector and the implicit latent feature vector to complement each other to get the user's latent feature vector

and the latent feature vector of the item

In step 203, according to the latent feature vector of the user and the latent feature vector of the item, the shallow linear preference feature and the deep nonlinear feature of the m users for the n items are determined.

Feed the latent feature vector of the user and the latent feature vector of the item into the hybrid model layer.

In the embodiment of the present invention, a dot product calculation is performed on the potential feature vector of the user and the potential feature vector of the item, as the input of the preset matrix decomposition model, and the matrix decomposition model outputs the shallow linear preference of m users for n items Features; take the user's latent feature vector and the item's latent feature vector as the input of the preset deep neural network model, and use the hidden layer of the multilayer perceptron of the deep neural network model to obtain the deep nonlinearity of m users for n items Preference feature, wherein the matrix decomposition and deep neural network joint prediction model includes matrix decomposition model and deep neural network model.

In the example of the present invention, the matrix factorization model and the deep neural network model are located in the hybrid model layer of the joint prediction model.

In the embodiment of the present invention, the dot product calculation is performed on the potential feature vector of the user and the potential feature vector of the item by formula 1, and formula 1 is:

Among them, the symbol refers to the product of the corresponding elements of the two vectors, that is, the dot product.

作为矩阵分解模型的输入。will be calculated

as input to the matrix factorization model.

In the embodiment of the present invention, the hidden layer of the multi-layer perceptron of the deep neural network model is used to obtain the deep nonlinear preference feature of the user for the item, and the complex relationship between the user and the item is modeled; the interaction between the user and the item is obtained. The multi-layer nonlinear projection of , the definition of multi-layer complex user preference features in the deep learning matrix factorization model layer is as follows: Equation 2:

...

和

为多层感知机模型第X层的权重矩阵、偏置向量和激活函数，z_EI为模型最初输入随机特征数据，

为第X-1层的输入特征数据，

为模型每层输出结果。in

and

is the weight matrix, bias vector and activation function of the Xth layer of the multilayer perceptron model, z _EI is the initial input random feature data of the model,

is the input feature data of the X-1 layer,

Output results for each layer of the model.

In step 204, a prediction result is obtained through a preset activation function according to the shallow linear preference feature and the deep nonlinear preference feature of m users for n items.

The feature vector corresponding to the user's shallow linear preference feature for the item and the feature vector corresponding to the deep nonlinear preference feature are connected and combined to predict the user's preference for the item.

用户u为m个用户中的任一用户，项目i为n个项目中的任一项目。如下公式3：In a possible implementation, in the prediction layer of the joint prediction model, the sigmoid function is used as the activation function of the prediction layer, and the shallow linear preference feature and the deep nonlinear preference feature of the user to the item are combined to predict the user u's preference for the item i. Preferences

User u is any one of m users, and item i is any one of n items. Formula 3 below:

为用户对项目的浅层线性偏好特征所对应的特征向量，

为用户对项目的深层非线性偏好特征所对应的特征向量。In the formula, a _out is the activation function of the prediction layer, h ^T is the weight parameter of the prediction layer,

is the feature vector corresponding to the user's shallow linear preference feature for the item,

is the feature vector corresponding to the user's deep nonlinear preference feature for the item.

For any user u, according to the user u's preference value for each item, the pre-set number of items are recommended for the user in the order of preference value from high to low, so as to improve the user's satisfaction.

In one possible implementation, the loss function is a crucial component in the construction of the predictive model, which is related to the performance of the recommendation algorithm. Its essence is the objective optimization function of the predictive model, which can be defined according to the explicit feedback data and the implicit feedback data of the user's interaction with the item. At present, the commonly used pointwise loss functions in recommendation algorithms are mainly square loss function and binary cross entropy loss function.

The squared loss function is often used in matrix decomposition recommendation algorithms based on explicit feedback data. Its basic definition is shown in Equation 4 below. Equation 4 is:

用于表示预测模型的输出值。In Equation 4, L _squ is used to represent the loss function, U is the user set, I is the item set, w _ui is used to represent the weight parameter of the prediction model, y _ui is used to represent the target value of the prediction model,

Used to represent the output value of the predictive model.

The binary cross-entropy loss function is often used in the deep learning recommendation model based on implicit feedback. Its basic definition is shown in Equation 5 below. Equation 5 is:

用于表示预测模型的输出值。In Equation 5, L _log is used to represent the loss function, U is the user set, I is the item set, y _ui is used to represent the target value of the prediction model,

Used to represent the output value of the predictive model.

In the embodiment of the present invention, the loss function of the joint prediction model of matrix decomposition and deep neural network is Equation 6, and Equation 6 is

L= _{ηL I} +(1-η)LE

Among them, L is the loss function of the joint prediction model of matrix decomposition and deep neural network, L _E is the loss function of explicit scoring, L _I is the loss function of implicit scoring, η is a preset coefficient, and η is greater than 0 and less than 1.

In the embodiment of the present invention, the explicit score is normalized to a value between 0 and 1, that is, the explicit score of all users for an item is divided by the maximum score value, for example, the score of user u for item i is 3 , the maximum score value is 5, then the explicit score normalized by user u to item i is 3÷5=0.6. At this time, the normalized display score is more suitable for the binary cross-entropy loss function. Based on this, in the embodiment of the present invention, the loss function of the explicit score is as follows in Formula 7, where Formula 7 is:

为矩阵分解与深度神经网络联合预测模型的预测值，即联合预测模型输出的用户u对项目i的偏好的预测值；Among them, L _E is the loss function of explicit scoring, UE is the number of users in the explicit scoring matrix, _IE is the number of items in the explicit scoring matrix, and r _ui is the actual value of the explicit scoring of item _i by user u , Max(R) is the maximum value of the explicit score,

is the predicted value of the joint prediction model of matrix factorization and deep neural network, that is, the predicted value of user u's preference for item i output by the joint prediction model;

In the embodiment of the present invention, the loss function of the implicit score is as shown in formula 8, and formula 8 is

为矩阵分解与深度神经网络联合预测模型的预测值，即联合预测模型输出的用户u对项目i的偏好的预测值。Among them, L _I is the loss function of implicit scoring, U _I is the number of users in the implicit scoring matrix, I _I is the number of items in the implicit scoring matrix, and ir _ui is the implicit scoring value of user u for item i ,

is the predicted value of the joint prediction model of matrix factorization and deep neural network, that is, the predicted value of user u's preference for item i output by the joint prediction model.

Most of the existing recommendation systems use error loss evaluation methods, such as: root mean square error (RMSE) and mean absolute error (MAE) to evaluate the similarity of users' predicted preferences and real preferences; but the method provided by the implementation of the present invention , is a list of the top N most interesting items recommended for users according to the user's ranking of item preference predictions. Compared with the evaluation method of error loss, the method of ranking evaluation is more practical.

In order to evaluate the effectiveness of the prediction method provided by the implementation of the present invention, the parameters of the joint prediction model of matrix decomposition and deep neural network are optimized and updated. Evaluation indicators, HR and NDCG are commonly used evaluation indicators of the recommendation system, which will not be repeated in the implementation of the present invention.

The embodiment of the present invention provides a new loss function, which combines the two evaluation indicators of HR and NDCG, and after multiple iterations, optimizes the parameters of the joint prediction model of matrix decomposition and deep neural network, and predicts users according to the optimized model parameters. For the preference value of the item, sort the preference value in descending order, and recommend the top N items in the sorting result to the user.

The embodiment of the present invention mines the user's shallow linear preference feature through a matrix decomposition model, and mines the user's deep nonlinear preference feature through a deep neural network model, and combines the two features to predict the user preference, thereby improving the accuracy of the user preference prediction. On the other hand, the embodiment of the present invention also provides a new loss function, which combines the evaluation index to optimize the parameters of the matrix decomposition and the deep neural network joint prediction model, so as to further improve the accuracy of the model for user preference prediction.

It should be understood that the size of the sequence numbers of the steps in the above embodiments does not mean the sequence of execution, and the execution sequence of each process should be determined by its functions and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

The following are apparatus embodiments of the present invention, and for details that are not described in detail, reference may be made to the above-mentioned corresponding method embodiments.

FIG. 3 shows a schematic structural diagram of an apparatus for predicting user preference provided by an embodiment of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, and the details are as follows:

As shown in FIG. 3 , the user preference prediction device 3 includes: a construction module 31 and a prediction module 32;

The building module 31 is used to construct an explicit scoring matrix and an implicit scoring matrix according to the historical interaction data between m users and n items. The explicit scoring matrix is used to represent the explicit scoring of m users on n items, and the implicit scoring matrix The formula rating matrix is used to represent the implicit rating of m users on n items. Both the explicit rating matrix and the implicit rating matrix are matrices with m rows and n columns;

The prediction module 32 is used to input the explicit scoring matrix and the implicit scoring matrix into a preset matrix decomposition and a deep neural network joint prediction model to obtain a prediction result. For each user in the m users, the prediction result includes the user's pair of The value of the degree of preference for each of the n items.

In the embodiment of the present invention, on the one hand, an explicit scoring matrix and an implicit scoring matrix are constructed in combination with the explicit scoring and implicit scoring of the item by the user, and the explicit scoring matrix and the implicit scoring matrix are used as the input of the prediction model , make full use of the respective features of explicit scoring and implicit scoring and the complementarity of the combination of the two to make it suitable for scoring scenarios. Combining the advantages of the two models further improves the accuracy of user preference prediction. Therefore, the method provided by the implementation of the present invention can improve the accuracy of the recommendation system in predicting the user's preference for an item, and display a personalized item recommendation list for the user. In order to provide users with more accurate personalized services, improve user satisfaction.

In a possible implementation manner, the prediction module 32 is configured to respectively determine the potential feature vector of the user and the potential feature vector of the item according to the explicit scoring matrix and the implicit scoring matrix;

According to the latent feature vector of users and the latent feature vector of items, determine the shallow linear preference feature and deep nonlinear preference feature of m users for n items;

According to the shallow linear preference feature and deep nonlinear preference feature of m users for n items, the prediction result is obtained through the preset activation function.

In a possible implementation manner, the prediction module 32 is configured to obtain the explicit latent feature vector and implicit latent feature vector of the user, and the explicit latent feature vector of the item through normal random initialization according to the displayed rating matrix and the implicit rating matrix. latent eigenvectors and implicit latent eigenvectors;

Add the user's explicit latent feature vector and the implicit latent feature vector to obtain the user's latent feature vector;

The item's latent eigenvector is obtained by summing the item's explicit latent eigenvector and its implicit latent eigenvector.

In a possible implementation, the prediction module 32 is configured to perform dot product calculation on the potential feature vector of the user and the potential feature vector of the item, as the input of the preset matrix factorization model, and the matrix factorization model outputs m user pairs of n Shallow linear preference features of each item;

The user's latent feature vector and the item's latent feature vector are used as the input of the preset deep neural network model, and the hidden layer of the multilayer perceptron of the deep neural network model is used to obtain the deep nonlinear preference features of m users for n items. , among which, the joint prediction model of matrix decomposition and deep neural network includes matrix decomposition model and deep neural network model.

In a possible implementation, the loss function of the joint prediction model of matrix factorization and deep neural network is

L= _{ηL I} +(1-η)LE

Among them, L is the loss function of the joint prediction model of matrix decomposition and deep neural network, L _E is the loss function of explicit scoring, L _I is the loss function of implicit scoring, η is a preset coefficient, and η is greater than 0 and less than 1.

In one possible implementation, the loss function showing the score is

为矩阵分解与深度神经网络联合预测模型的预测值；Among them, L _E is the loss function of explicit scoring, UE is the number of users in the explicit scoring matrix, _IE is the number of items in the explicit scoring matrix, and r _ui is the actual value of the explicit scoring of item _i by user u , Max(R) is the maximum value of the explicit score,

is the predicted value of the joint prediction model of matrix factorization and deep neural network;

The loss function for implicit scoring is

为矩阵分解与深度神经网络联合预测模型的预测值。Among them, L _I is the loss function of implicit scoring, U _I is the number of users in the implicit scoring matrix, I _I is the number of items in the implicit scoring matrix, and ir _ui is the implicit scoring value of user u for item i ,

is the prediction value of the combined prediction model of matrix factorization and deep neural network.

In a possible implementation manner, the building module 31 is used to remove abnormal data in the initial data set to obtain a target data set, wherein the initial data set includes historical interaction data of multiple users and multiple items, and for each user, If the number of historical interactions between the user and multiple items in the initial data set is less than or equal to the first preset number of times, the user's historical interaction data is abnormal data, and, for each item, if multiple users in the initial data set interact with the user If the historical interaction times of an item is less than or equal to the second preset number of times, the historical interaction data of the item is abnormal data, the first preset number of times is the same or different from the second preset number of times, and the target data set includes m users and n items historical interaction data.

In a possible implementation manner, the building module 31 is configured to normalize the score of the user u on the item i if the user u has a score for the item i, and determine the normalized score value. If item i has no score, set the score of user u to item i to 0, and construct an explicit score matrix according to the score of each user for each item. For any normalized score value, the score value is greater than equal to 0 and less than or equal to 1;

If user u interacts with item i, set user u's implicit score to item i to 1; if user u has no interaction with item i, set user u's implicit score to item i to 0, according to The implicit rating of each user for each item, constructing an implicit rating matrix.

The device for predicting user preference provided in this embodiment can be used to execute the above embodiments of the method for predicting user preference, and its implementation principle and technical effect are similar, and details are not described herein again in this embodiment.

FIG. 4 is a schematic diagram of a terminal provided by an embodiment of the present invention. As shown in FIG. 4 , the terminal 4 of this embodiment includes: a processor 40 , a memory 41 , and a computer program 42 stored in the memory 41 and executable on the processor 40 . When the processor 40 executes the computer program 42 , the steps in each of the foregoing embodiments of the user preference prediction method are implemented, for example, steps 101 to 102 shown in FIG. 1 . Alternatively, when the processor 40 executes the computer program 42, the functions of the modules/units in each of the foregoing apparatus embodiments, such as the functions of the units 31 to 32 shown in FIG. 3, are implemented.

Exemplarily, the computer program 42 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 41 and executed by the processor 40 to complete the this invention. The one or more modules/units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program 42 in the terminal 4 .

The terminal 4 may be a computing device such as a desktop computer, a notebook, a palmtop computer and a cloud server. The terminal 4 may include, but is not limited to, a processor 40 and a memory 41 . Those skilled in the art can understand that FIG. 4 is only an example of the terminal 4, and does not constitute a limitation to the terminal 4. It may include more or less components than the one shown in the figure, or combine some components, or different components, such as The terminal may also include input and output devices, network access devices, buses, and the like.

The processor 40 may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processors, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory 41 may be an internal storage unit of the terminal 4 , such as a hard disk or a memory of the terminal 4 . The memory 41 may also be an external storage device of the terminal 4, such as a plug-in hard disk equipped on the terminal 4, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, Flash card (Flash Card) and so on. Further, the memory 41 may also include both an internal storage unit of the terminal 4 and an external storage device. The memory 41 is used to store the computer program and other programs and data required by the terminal. The memory 41 can also be used to temporarily store data that has been output or will be output.

Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. Module completion, that is, dividing the internal structure of the device into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment may be integrated in one processing unit, or each unit may exist physically alone, or two or more units may be integrated in one unit, and the above-mentioned integrated units may adopt hardware. It can also be realized in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing from each other, and are not used to limit the protection scope of the present application. For the specific working processes of the units and modules in the above-mentioned system, reference may be made to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

In the foregoing embodiments, the description of each embodiment has its own emphasis. For parts that are not described or described in detail in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of the present invention.

In the embodiments provided by the present invention, it should be understood that the disclosed apparatus/terminal and method may be implemented in other manners. For example, the device/terminal embodiments described above are only illustrative. For example, the division of the modules or units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or Components may be combined or may be integrated into another system, or some features may be omitted, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated modules/units, if implemented in the form of software functional units and sold or used as independent products, may be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments, and can also be completed by instructing relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed by the processor, the steps of each of the foregoing embodiments of the user preference prediction method can be implemented. Wherein, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate form, and the like. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable media may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer-readable media Excluded are electrical carrier signals and telecommunication signals.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to implement the foregoing implementations. The technical solutions described in the examples are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the within the protection scope of the present invention.

Claims (10)

1. a user preference prediction method, is characterized in that, comprises: According to the historical interaction data of m users and n items, an explicit scoring matrix and an implicit scoring matrix are constructed. The explicit scoring matrix is used to represent the explicit scoring of the n items by the m users. The implicit scoring matrix is used to represent the implicit scoring of the n items by the m users, and both the explicit scoring matrix and the implicit scoring matrix are matrices with m rows and n columns; Inputting the explicit scoring matrix and the implicit scoring matrix into a preset matrix decomposition and a deep neural network joint prediction model to obtain a prediction result, for each of the m users, the prediction result includes the The value of the user's preference for each of the n items. 2. The method according to claim 1, wherein the inputting the explicit scoring matrix and the implicit scoring matrix into a preset matrix decomposition and a deep neural network joint prediction model, and obtaining the prediction result comprises: According to the explicit scoring matrix and the implicit scoring matrix, determine the potential feature vector of the user and the potential feature vector of the item, respectively; According to the latent feature vector of the user and the latent feature vector of the item, determine the shallow linear preference feature and the deep nonlinear preference feature of the m users for the n items; According to the shallow linear preference feature and the deep nonlinear preference feature of the m users for the n items, the prediction result is obtained through a preset activation function. 3. The method according to claim 2, wherein, according to the explicit scoring matrix and the implicit scoring matrix, respectively determining the potential feature vector of the user and the potential feature vector of the item comprises: : According to the displayed rating matrix and the implicit rating matrix, through normal random initialization, the explicit latent eigenvector and the implicit latent eigenvector of the user, and the explicit latent eigenvector and the implicit latent eigenvector of the item are obtained; adding the user's explicit latent feature vector and the implicit latent feature vector to obtain the user's latent feature vector; The item's explicit latent feature vector and the implicit latent feature vector are added to obtain the item's latent feature vector. 4. The method according to claim 3, wherein the shallow linearity of the m users to the n items is determined according to the latent feature vector of the user and the latent feature vector of the item Preference features and deep nonlinear preference features include: The dot product calculation is performed on the potential feature vector of the user and the potential feature vector of the item, as the input of a preset matrix factorization model, and the matrix factorization model outputs the shallow results of the m users on the n items. Layer linear preference feature; The potential feature vector of the user and the potential feature vector of the item are used as the input of the preset deep neural network model, and the hidden layer of the multi-layer perceptron of the deep neural network model is used to obtain the information about the m users. The deep nonlinear preference feature of the n items, wherein the matrix decomposition and deep neural network joint prediction model includes the matrix decomposition model and the deep neural network model. 5. The method according to any one of claims 1 to 4, wherein the loss function of the matrix decomposition and the deep neural network joint prediction model is L= _{ηL I} +(1-η)LE Wherein, L is the loss function of the matrix decomposition and the deep neural network joint prediction model, L _E is the loss function of the explicit score, L _I is the loss function of the implicit score, n is the preset coefficient, n is greater than 0 and less than 1. 6. The method according to claim 5, wherein the loss function of the display score is

Wherein, L _E is the loss function of the explicit scoring, UE is the number of users in the explicit scoring matrix, _IE is the number of items in the explicit scoring matrix, and r _ui is the user u to item _i The actual value of the explicit score, Max(R) is the maximum value of the explicit score,

is the predicted value of the combined prediction model of the matrix decomposition and the deep neural network; The loss function of the implicit score is

Among them, L _I is the loss function of implicit scoring, U _I is the number of users in the implicit scoring matrix, I _I is the number of items in the implicit scoring matrix, ir _ui is the implicit score of user u on item i rating value,

is the prediction value of the combined prediction model of matrix factorization and deep neural network. 7. The method according to any one of claims 1 to 4, characterized in that, before constructing an explicit scoring matrix and an implicit scoring matrix according to the historical interaction data of m users and n items, the method further comprises: : Remove the abnormal data in the initial data set to obtain the target data set, wherein the initial data set includes historical interaction data of multiple users and multiple items, for each user, if the user and multiple items in the initial data set The number of historical interactions is less than or equal to the first preset number of times, then the user's historical interaction data is abnormal data, and, for each item, if the number of historical interactions between multiple users in the initial data set and the item is less than or equal to the first Two preset times, the historical interaction data of the item is abnormal data, the first preset times and the second preset times are the same or different, and the target data set includes the m users and n items historical interaction data. 8. The method according to claim 7, wherein, according to the historical interaction data of m users and n items, constructing an explicit scoring matrix and an implicit scoring matrix comprises: If user u scores item i, then normalize user u's score for item i to determine the normalized score value; if user u does not score item i, then use user u's score for item i Set to 0, construct the explicit score matrix according to the score of each user for each item, and for any normalized score value, the score value is greater than or equal to 0 and less than or equal to 1; If user u interacts with item i, set user u's implicit score to item i to 1; if user u has no interaction with item i, set user u's implicit score to item i to 0, according to The implicit rating of each user for each item, constructing the implicit rating matrix. 9. A terminal comprising a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor implements the above claims when executing the computer program The steps of any one of 1 to 8. 10. A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 8 when the computer program is executed by a processor A step of.

Images