CN120013637A

CN120013637A - A method and device for intelligent recommendation with coordinated accuracy and diversity

Info

Publication number: CN120013637A
Application number: CN202510074562.9A
Authority: CN
Inventors: 赵小敏; 李欣蔚
Original assignee: Zhejiang University of Technology ZJUT
Current assignee: Zhejiang University of Technology ZJUT
Priority date: 2025-01-17
Filing date: 2025-01-17
Publication date: 2025-05-16
Anticipated expiration: 2045-01-17
Also published as: CN120013637B

Abstract

A smart recommendation method for coordinated diversity and accuracy, comprising: loading a dataset containing user-item interaction records; initializing user and item embeddings, assigning a randomly initialized embedding vector with a dimension of d to each user and item; constructing a bipartite graph of users and items, in which users and items are nodes and interaction relationships are edges; using a category-aware neighbor selection strategy to screen the neighbors of the nodes and generate a category-balanced neighbor subset; using a lightweight graph convolutional network to aggregate information on the category-balanced neighbor subset, using a hierarchical dynamic attention module to weightedly fuse multiple layers of embedding, balance information at different levels, and update user and item embedding vectors; adopting a multi-task joint training strategy, adjusting the Bayesian personalized loss based on adaptive fusion weights, and combining the main task Bayesian personalized loss and auxiliary task contrast learning for training and optimization; based on the final trained user embedding and item embedding, calculating the inner product between them to obtain the user's rating of the item, and presenting the recommendation list to the user.

Description

Intelligent recommendation method and device with cooperative accuracy and diversity

Technical Field

The invention belongs to the technical field of artificial intelligence, in particular relates to an intelligent recommendation method and device for optimizing trade-off of accuracy and diversity in a recommendation system, and aims to improve comprehensiveness and accuracy of user experience.

Background

In the age of information flooding today, the amount of data newly added every day increases exponentially, so that the acquisition and digestion of information by the masses become more difficult, users often go no way down when facing massive amounts of information, and screening out truly valuable information becomes a great challenge. In this context, recommendation systems have been developed as an important tool for alleviating information overload. By analyzing the behaviors and preferences of the user, the recommendation system can efficiently provide relevant content for the user, so that information acquisition efficiency and user experience are improved.

Generally, evaluating the effectiveness of a recommender system generally takes accuracy as a vital indicator to measure the likelihood of a user interacting with a particular item. However, accurate recommendations are not necessarily satisfactory, and merely optimizing the accuracy of the recommendation system may exacerbate the 'filtered bubbles' effect, i.e., the user is limited to a known range of interest, has difficulty in accessing novel, diverse content, and weakens the exploratory for new points of interest. For example, when purchasing clothes on an e-commerce platform such as a Taobao, a concert, etc., a user spends a great deal of time browsing autumn and winter clothes and expecting to find a new wearing style, if a great deal of accurate and similar clothes are recommended, the user may be tired aesthetically, reducing the user's desire to purchase and the platform viscosity. In recent years, research shows that improving the diversity of recommended content can obviously improve core service indexes such as clicking, residence time and long-term retention rate of users. By increasing the difference between recommended items, the diversified recommendation system can better capture and meet different interests of users, help the users find potential interests, enrich user experience and create greater value for the platform. However, optimizing diversity alone tends to result in reduced accuracy of the recommendation, and thus how to balance accuracy and diversity, with minimal accuracy cost for diversity is a direction of interest.

By representing the user's historical interactions as a user-project bipartite graph, the graph-based approach is able to efficiently capture high-level connection information. The Graph Neural Network (GNNs) is used as a powerful learning method for processing graph structure data, is widely applied to graph-based recommendation systems, and a typical graph-based recommendation system aggregates information of each node neighborhood to generate node embedding by designing a proper graph structure and a neural network, so that new possibility is provided for diversified recommendation. However, in graph-based diversified recommendation, firstly, the problem of overcomplete is easily caused by directly stacking each layer of information, so that the recommendation accuracy is reduced, and secondly, how to effectively control the neighborhood to increase the diversity is also a problem, if all neighbors are directly aggregated, long tail items are submerged by popular items, and the multi-dimensional interests of users cannot be captured. In general, the research of the conventional diversified recommendation method based on the graph neural network is still relatively limited, and the dilemma between the accuracy and the diversity of the trade-off is often faced.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, provides an intelligent recommendation method and device with cooperative precision and diversity, and aims to solve the contradiction between diversity and precision in the prior art.

The method combines static and dynamic attentiveness to adaptively allocate the importance of each layer of information, enhance the capability of capturing local and high-order structural information by the model, alleviate the overcomplete problem existing in deep GNN, 3. Cross-view information fusion weight strategy, grasp the essential attribute of the object from multiple view angles by fusing the category information and the value information of the object, dynamically adjust the importance of different objects, realize finer balance of popular object and long-tail object, 4. Long-tail oriented contrast learning optimization training module, integrate contrast learning into the frame by dynamic sampling and noise disturbance embedded in space, enhance the representation learning capability of the long-tail object, and promote the diversified learning effect. Overall, this approach achieves an excellent tradeoff of recommendation accuracy and diversity.

The first aspect of the invention relates to an intelligent recommendation method with cooperative precision and diversity, which comprises the following steps:

s1, loading a data set containing interaction records of a user and an article;

s2, initializing user and article embedding, and distributing a randomly initialized embedding vector for each user and article, wherein the dimension is d;

s3, constructing a bipartite graph of the user and the article, wherein the user and the article are used as nodes, and the interaction relationship is used as an edge;

s4, screening neighbors of the nodes by using a category-aware neighbor selection strategy to generate a category-balanced neighbor subset;

S5, information aggregation is carried out on neighbor subsets with balanced categories by using a lightweight graph rolling network, multi-layer embedding is carried out by using a layered dynamic attention module for weighting fusion, information of different layers is balanced, and user and article embedding vectors are updated;

S6, adopting a multi-task combined training strategy, adjusting Bayesian personalized loss based on self-adaptive fusion weight, and simultaneously combining main task Bayesian personalized loss and auxiliary task comparison learning to train and optimize;

and S7, calculating the inner product between the user and the article based on the finally trained user embedding and article embedding to obtain the scores of the user on the articles, and selecting K articles with the highest scores as recommendation lists to be presented to the user.

Further, step S4 includes:

The method aims at improving embedding representativeness and diversity in a recommendation system by optimizing a neighbor selection process. The traditional sub-nearest neighbor selection method generally uses a greedy algorithm to maximize the maximum entropy function, the algorithm starts from an empty set S _u, i epsilon N _u\S_u with maximized marginal gain is selected each time, and a diversity nearest neighbor subset is obtained after k steps of greedy selection. This presents a implicit problem in that when categories are severely unbalanced, if greedy strategies are chosen for maximum similarity and representativeness, the model will tend to choose those hot categories that cover more nodes on the similarity measure, since these hot categories occupy the majority of the entire neighborhood Nu, and cold items tend to be ignored relatively speaking because of the smaller marginal gain, which is particularly pronounced when k is smaller. In order to solve the above problem, the method optimizes the original neighbor searching process, introduces a class equalization penalty term to adjust the original maximum entropy function, so that the distribution S _u of the selected subset class is as close as possible to the distribution of the neighborhood N _u, and the final function is defined as follows:

The higher the function value of the adjusted maximum entropy function f (S _u), the more effectively the selected subset can represent the whole neighbor set, and the diversity and the balance of the categories can be maintained. Wherein the method comprises the steps of A penalty term is expressed that reduces the difference between the proportion of the current category in the selected set and the target proportion, and when a popular category node is selected too much, the penalty term increases, weakening the entropy gain of that category, thereby reducing its probability of selection in the next step, and reducing its probability of selection in the next step. sim (i, i ^′) represents the similarity between the two terms, and in the method, the mahalanobis distance is used for measuring the similarity between samples, so that after k steps of selection, a diversified and balanced neighbor subset of each user is obtained and used for subsequent information aggregation operation.

Further, step S5 includes:

S51, adopting a lightweight graph convolution LGC as a basic GNN layer, directly aggregating and diversifying information of neighbor nodes, updating embedded representations of users and articles in different layers, wherein a specific embedded updating formula is as follows:

Wherein S _u and S _i respectively represent neighbor sets of a user u and an article i obtained through category-aware neighbor selection; The normalization factor is used for avoiding overlarge embedded characterization values caused by multiple aggregation operations, controllable uniform disturbance is added in the layer-by-layer information aggregation process, and the generated embedded is stored as an intermediate layer and used as a subsequent comparison learning view.

S52, in the graph neural network, different layers generate embedments by aggregating information from neighbor nodes with different hop counts, and the first layer mainly aggregates information from l-hop neighbors, and the hierarchical aggregation provides opportunities for introducing diversified information from low-order neighbors and high-order neighbors. However, conventional layer attention mechanisms typically rely on static weight assignments, which makes it difficult to adaptively capture Gao Jielin centered diversity information, over-rely on low-order neighbors, and further limit the diversity of the recommendation. In order to overcome the limitation, the method provides a multi-level dynamic perception attention module, comprehensively considers dynamic and static attention, and considers the personalized characteristics and the overall stability of the node from the global and local perspectives. The static attention weight is calculated based on a common layer attention mechanism and is used for measuring the importance of embedding of each layer, and the parameter W _Att∈R^d of the attention weight is set, wherein the specific calculation formula is as follows:

The weight calculation of the dynamic attention mechanism is based on the initial embedding e ⁽⁰⁾ and each layer embedding e ^(l) of the node, and the hierarchical bias b ^(l) is combined to capture the personalized characteristics of the node, and the design can reflect the core characteristics of the node more accurately without being interfered by neighbor information. The specific formula is as follows:

Where the function g (x, y, z) = < x, y > +z, the personalization information and the hierarchy information are combined by inner product, and the perceptibility between layers is enhanced by the hierarchy bias. The final attention weight assignment is determined by both static and dynamic weights, and the final embedding of node u (node i is also the same) is expressed as:

The dynamic attention mechanism effectively captures the personalized needs of the nodes, while the static attention mechanism ensures the global stability among layers. Compared with the traditional method, the method effectively relieves the problem that the traditional layer attention mechanism excessively depends on low-order neighbors, improves diversity and information coverage range, and simultaneously relieves the problem of excessive smoothness in deep GNN.

Further, step S6 includes:

S61, fusing article value information and category information on the premise of keeping category weighting and improving long-tail attention, further capturing essential information of the fused article, and calculating final weights of the article fusion value information and the category information;

and S62, adjusting the loss calculation of the main task according to the weight of the article after fusing the value and the category information, guiding and comparing the sample sampling of the learning task based on the weight, and finally carrying out joint optimization training on the main task and the auxiliary task.

Still further, step S61 includes:

Optimizing the average loss of all samples in the recommendation system may ignore long tail categories, resulting in insufficient diversity of recommendation results. To alleviate this problem, conventional strategies typically calculate weights based on categories, assign lower weights to trending categories, assign higher weights to trending categories, and re-weight computational losses, thereby increasing the attention of the model to items of the trending categories, and increasing diversity. However, this "one-shot" approach ignores the difference in items within the category that the existence of otherwise valuable good items in the hot category should not be stressed on a contemporaneous basis, while uneven quality in the cold category, too general computational weighting, may result in unnecessary noise. Therefore, the method provides a weight adjustment strategy for fusing the article type and the value information, and the article value factor v (i) is introduced to further distinguish the similar articles on the premise of improving the long tail attention degree by keeping the type weighting. The model can be helped to jump out of the limitation of the item category information view angle by combining the value information view angle of the item, and further the essential information of the fusion item is captured, so that the accuracy and the diversity are more reasonably weighted. The item value factor v (i), which may be measured by the frequency of user interactions of the item, such as click-through rate, number of purchases, etc. The final weight W _opt (i) has the following formula:

Wherein the method comprises the steps of The method is normalized class weight, reflects the scarcity of the class, and is higher as the number of class samples is smaller, v (i) is the interactive value of the object i, such as normalized user click times or purchase times, and alpha is an adjusting parameter used for controlling the influence of the object interactive value on the final weight.

Still further, step S62 includes:

Long tail items are often difficult to adequately capture because of the sparse number of user interactions, resulting in model preference and focusing phenomena on the flowing items. Aiming at the problem, the method introduces a contrast learning task based on the optimization of the main task BPR to improve the uniformity and the robustness of the embedded distribution, thereby enhancing the representation and the capturing capability of the model on long-tail objects. The core idea is to combine a dynamic sampling cooling mechanism to sample subsets on the basis of obtaining class balance weight W _opt (i), and focus on cold class or long-tail objects through the contrast learning of uniform noise disturbance, so that the generalization capability and recommendation diversity of the model are improved. Specifically, items are weighted with weights W _opt (i) to promote the probability of entry of long-tailed and higher value items when the subset is sampled, while setting a dynamic sampling upper limit for each item. When the sampling frequency of an article in the comparison learning subset reaches the upper limit, the method can reduce the subsequent probability of entering the article or temporarily reject the article from the candidate set. Through the current limiting mechanism, excessive attention given to individual long-tail high-value articles by contrast learning can be prevented, proper global balance of subsets is ensured to be maintained when diversity is improved, and then finer-granularity hierarchical information alignment is realized by means of uniform disturbance noise and cross-layer contrast learning encouragement models. The optimized joint loss function is as follows:

In the method of the invention, not only is the characterization z _i of the final layer focused on for a given user-article pair, but the intermediate layer is also embedded Incorporating the contrast target enables model embedding in the representation space to capture more stereoscopic features by implementing contrast learning between multiple levels of embedded representations. And finally, when the loss is calculated, combining a main task (BPR) and Contrast Learning (CL), wherein the main task loss is focused on optimizing the preference sequence of positive and negative samples of a user so as to improve the recommendation accuracy, and the contrast learning loss aims at enhancing the diversity and the robustness of the representation space and improving the capturing capability of long-tail articles. And mu is a balance parameter and is used for controlling the contribution degree of the contrast learning loss in the whole optimization process, and when mu is increased, the influence of contrast learning is increased, so that the representation capability and diversity of long-tail articles are further promoted. Through the combined training mode, the model can be combined with multi-level embedded multi-view signals, so that good performance can be kept on the recommended precision level, and diversity and robustness can be effectively improved.

The second aspect of the invention relates to an intelligent recommendation device of cooperative diversity and accuracy, comprising a memory and one or more processors, wherein executable codes are stored in the memory, and the one or more processors are used for realizing the intelligent recommendation method of cooperative diversity and accuracy of the invention when executing the executable codes.

A third aspect of the present invention relates to a computer readable storage medium having stored thereon a program which, when executed by a processor, implements the intelligent recommendation method of collaborative diversity and accuracy of the present invention.

The method has the innovation points that a class-aware neighbor selection strategy is introduced, equalization and diversification of neighborhood classes are realized through dynamic adjustment of a maximum entropy function, a hierarchical dynamic perception attention mechanism is adopted, information weights of all layers are distributed in a self-adaptive mode, the problem of excessive smoothness of a deep model is effectively solved, a cross-view information fusion strategy is provided by combining object class and value information to optimize weight distribution, a long-tail oriented contrast learning optimization module is designed, and the representation capacity and the diversity learning effect of long-tail objects are enhanced. Through the multi-task combined optimization training, the method breaks through the trade-off limitation of the traditional algorithm between accuracy and diversity, and greatly improves the recommendation quality and user experience.

The method has the advantages that a category-aware neighbor selection strategy, a hierarchical dynamic attention mechanism and a cross-view information fusion strategy are introduced, the trade-off of accuracy and diversity in a recommendation system is optimized, the diversity and individuation of recommended contents are remarkably improved, the Bayesian individuation sequencing is combined with auxiliary tasks comprising long-tail guide comparison learning optimization modules by adopting a multi-task combined training strategy, the representation capacity of long-tail projects and the generalization of models are enhanced, meanwhile, a lightweight graph convolution network is adopted, the calculation complexity is effectively reduced, the training efficiency of the models is improved, the content exploration experience and the platform viscosity of users are remarkably improved while the recommendation accuracy is ensured, and higher technical innovation and application value are presented.

Drawings

FIG. 1 is a schematic flow chart of the present invention;

FIG. 2 is a diagram of a model training framework of the present invention;

Fig. 3 is a schematic view of the apparatus of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

Example 1

As shown in fig. 1, an intelligent recommendation method with cooperative accuracy and diversity includes the following specific steps:

step S1, loading a data set containing user interaction records with the article

In some embodiments, the disclosed dataset may be obtained from a website such as Kaggle, which typically contains information such as user id, item id, timestamp, etc.;

Step S2, initializing user and article embedding

Based on the loaded data set, an initial embedded vector is generated for each user and item, with the vector dimension set to d. Vector initialization may use a random distribution or pre-trained model, for example, by matrix decomposition methods to generate initial values. These embedded vectors are used to represent features of the user and the item, which will be updated dynamically in response to training. The user-item ID is mapped to a look-up table E ⁽⁰⁾ of dense vectors, from which the corresponding embedded representation can be derived by indexing, the initial embedded look-up table for the user and item is as follows:

step S3, constructing a user-object bipartite graph

And extracting the relation between the user and the object from the interaction record of the user, and constructing a bipartite graph. The user node and the article node are respectively used as two types of nodes of the bipartite graph, the interaction behavior of the user and the article is used as an edge, and the weight of each edge can be set according to the interaction strength (such as the click times or the purchase frequency).

Step S4, screening neighbors of the nodes by using a category-aware neighbor selection strategy to generate a category-balanced neighbor subset

The invention uses a new class balance neighbor selection strategy to adjust the original neighbor searching process, introduces a class balance punishment item, enables the distribution S _u of the selected subset class to be as close as possible to the distribution of the neighborhood N _u, measures the diversity through the adjusted maximum entropy function, and defines the final function as follows:

The higher the value of f (S _u) function, the more the selected neighbor subset is represented, which not only can effectively represent the whole neighbor set, but also can maintain the diversity and the balance of the categories. Wherein the method comprises the steps of And expressing a penalty term for adjusting the difference between the proportion of the current category in the selected set and the target proportion, wherein when the nodes of the popular category are excessively selected, the penalty term is increased, the entropy gain of the category is weakened, the selection probability of the class in the next step is reduced, the distribution of the finally selected neighbor nodes is enabled to be as close to the initial neighborhood distribution as possible through the function design, and the diversified and balanced neighbor subsets of each user can be obtained through the selection after the adjustment of k steps.

S5, information aggregation is carried out on the neighbor subsets with balanced categories by using a lightweight graph rolling network, and multi-layer embedding is weighted and fused by using a layered dynamic attention module

The lightweight graph convolution LGC is adopted as a basic GNN layer, the information of diversified and balanced neighbor nodes is directly aggregated, the embedded representation of users and objects in different layers is updated, and a specific embedded updating formula is as follows:

wherein S _u and S _i respectively represent neighbor sets of the user u and the object i obtained through category perception selection, and the aggregation process is that of the category perception information in fig. 1, and meanwhile, uniformly controllable disturbance is added to the embedding to serve as an intermediate layer to be stored and used as a subsequent comparison learning view. In the graph neural network, different layers generate embeddings by aggregating information from neighbor nodes of different hop counts, and the first layer mainly aggregates information from l-hop neighbors. After the LGC obtains embedded representation of aggregation neighbor information of different layers, the method uses a dynamic and static combined attention mechanism, comprehensively considers global and local view angles, and takes the personalized characteristics and the overall stability of the node into consideration. The static attention weight is calculated based on a common layer attention mechanism to measure the importance of each layer of embedding. The dynamic attention mechanism adopts node initial embedding to assist in generating dynamic attention weight, so that the model can consider the essential characteristics and diversity targets of the nodes when multi-layer embedding is aggregated. The specific formula is as follows:

s6, adopting a multi-task combined training strategy, combining the Bayesian personalized loss of the main task and the contrast learning of the auxiliary task to train and optimize

As shown in fig. 2, aiming at the defects existing in category weight calculation, the method provides an adjustment strategy combining the category of the article and the value information, and introduces the value factor v (i) of the article to further divide the category of the article on the premise of keeping the category weight and improving the long tail attention. The item value factor may be calculated based on the frequency of user interactions (e.g., clicks, purchases) of the item, and the adjusted weight W _opt (i) may be calculated as:

And adjusting the loss calculation of the main task according to the weight of the article after fusing the value and the category information, and guiding the sample sampling of the comparison learning task based on the weight. In the sampling process, the probability of entry of an article is adjusted according to the weight W _opt (i), meanwhile, a dynamic sampling upper limit is set for each article, and when the sampling frequency of an article in the comparison learning subset reaches the upper limit, the subsequent probability of entry is reduced or the article is temporarily removed from the candidate set. Through the current limiting mechanism, excessive attention to individual long-tail high-value articles by contrast learning can be prevented, and proper global balance of subsets is ensured when diversity is improved. Through cross-layer contrast learning, the intermediate layer and the final embedded layer representation of the record retained in step S5 are presented as two different views of the contrast learning, encouraging the model to achieve finer granularity of hierarchical information alignment. And finally, carrying out joint optimization training on the main task and the auxiliary task. The specific loss function is designed as follows:

in the present method, not only is the characterization z _i of the final layer focused on for a given user-object pair, but the embedding of the intermediate layer Incorporating the contrast target enables model embedding in the representation space to capture more stereoscopic features by implementing contrast learning between multiple levels of embedded representations. And finally, when the loss is calculated, combining a main task (BPR) and Contrast Learning (CL), wherein the main task loss is focused on optimizing the preference sequence of positive and negative samples of a user so as to improve the recommendation accuracy, and the contrast learning loss aims at enhancing the diversity and the robustness of the representation space and improving the capturing capability of long-tail articles. Through the combined training mode, the model can be combined with multi-level embedded multi-view signals, so that good performance can be kept on the recommended precision level, and diversity and robustness can be effectively improved.

Step S7, calculating the inner product between the user and the object based on the finally trained user embedding and object embedding to obtain the scores of the user on the objects, and recommending the K objects with the highest scores to the user

Finally, the trained embedded representations e _u and e _i are reserved, the interaction probability between the items can be obtained through inner product calculation, the inner product result represents the relevance score of the user u and the item i, and the higher the score is, the greater the interest degree of the user in the item is. Specifically, for each user u, a set of matching degree scores is obtained by calculating the relevance scores of the user u and all candidate items, a recommendation result is output according to the descending order, K items with the highest scores are usually intercepted and output as a final recommendation result, and the output form of the recommendation result can be an item list or a specific commodity display page.

Example 2

Referring to fig. 3, the present embodiment relates to an intelligent recommendation apparatus for collaborative diversity and accuracy, which includes a memory and one or more processors, wherein executable codes are stored in the memory, and the one or more processors are used for implementing the intelligent recommendation method for collaborative diversity and accuracy of embodiment 1 when executing the executable codes.

Example 3

The present embodiment relates to a computer-readable storage medium having stored thereon a program which, when executed by a processor, implements the intelligent recommendation method of collaborative diversity and accuracy of embodiment 1.

It should be noted that, although the examples described above are illustrative, this is not a limitation of the present invention, and thus the present invention is not limited to the above-described specific embodiments. Other embodiments, which are apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein, are considered to be within the scope of the invention as claimed.

Claims

1. An intelligent recommendation method for coordinating diversity and accuracy, characterized by comprising the following steps:

S1: Load the dataset containing the interaction records between users and items;

S2: Initialize user and item embeddings, assigning a randomly initialized embedding vector of dimension d to each user and item;

S3: Construct a bipartite graph of users and items, where users and items are nodes and interaction relationships are edges;

S4: Use a category-aware neighbor selection strategy to screen the neighbors of the node and generate a category-balanced neighbor subset;

S5: Use a lightweight graph convolutional network to aggregate information on a category-balanced subset of neighbors, use a hierarchical dynamic attention module to weightedly fuse multiple layers of embedding, balance information at different levels, and update user and item embedding vectors;

S6: Adopting a multi-task joint training strategy, adjusting the Bayesian personalized loss based on adaptive fusion weights, and combining the main task Bayesian personalized loss and auxiliary task comparative learning for training optimization;

S7: Based on the final trained user embedding and item embedding, the inner product between them is calculated to obtain the user's rating of the item, and the K items with the highest ratings are selected as the recommendation list and presented to the user.

2. The method according to claim 1, characterized in that step S4 specifically comprises:

A maximum entropy function that takes into account the balanced distribution of categories is designed, so that the category distribution of the selected subset _Su is as close as possible to the distribution of the neighborhood _Nu . The diversity is measured by the adjusted maximum entropy function, which is defined as follows:

The higher the value of the function, the better the representativeness and category diversity of the selected neighborhood subset; Represents a penalty term, which is used to adjust the difference between the proportion of the current category in the selected set and the target proportion. When too many popular category nodes are selected, the penalty term increases, weakening the entropy gain of the category, thereby reducing its selection probability in the next step. After the neighbor selection strategy after k steps of adjustment, a diversified and balanced neighbor subset for each user is obtained for subsequent aggregation operations.

3. The method according to claim 1, characterized in that step S5 specifically comprises:

S51: Use lightweight graph convolution LGC as the basic GNN layer to directly aggregate the information of diverse and balanced neighbor nodes and update the embedding representation of users and items at different layers. The specific embedding update formula is as follows:

Where _Su and _Si represent the neighbor sets of user u and item i respectively obtained through module selection; It is a normalization factor, which is used to avoid multiple aggregation operations that lead to excessive embedding representation values. It adds controllable uniform perturbations in the process of layer-by-layer information aggregation. The generated embedding is saved as an intermediate layer and used as a subsequent contrastive learning view.

S52: Construct a multi-level dynamic perception attention module, combining dynamic attention and static attention, taking into account both the personalized characteristics and overall stability of nodes from a global and local perspective; static attention weights are calculated based on the ordinary layer attention mechanism to measure the importance of each layer embedding; the weight calculation of the dynamic degree attention mechanism is based on the initial embedding e ⁽⁰⁾ of the node and the embedding e ^(l) of each layer, combined with the layer bias b ^(l) to capture the personalized characteristics of the node; the initial embedding of the node is used to assist in generating the dynamic attention weight, the specific formula is:

The function g(x, y, z) = <x, y> + z combines personalized information and hierarchical information through inner products, and enhances the perception ability between multiple layers through hierarchical bias; the final attention weight distribution is determined by static and dynamic weights, and the final embedding of node u is expressed as:

By adjusting the fusion hyperparameter η, the model can capture focused global and local information while alleviating the over-smoothing problem in deep GNNs.

4. The method according to claim 1, characterized in that step S6 specifically comprises:

S61: Under the premise of retaining the category weighting to improve the long-tail attention, the value information of the item and the category information are integrated to further capture the essential information of the integrated item and calculate the final weight of the integrated value information and category information of the item;

S62: Adjust the loss calculation of the main task according to the weight of the fusion value and category information of the item, and guide the sample sampling of the comparative learning task based on the weight, and finally jointly optimize the main task and the auxiliary task.

5. The method according to claim 4, characterized in that step S61 specifically comprises:

The weighting strategy that integrates item category and value information, while retaining the category weighting to improve the long-tail attention, introduces the item value factor v(i) to further distinguish items; the item value factor is measured by the user interaction frequency, and the final weight W _opt (i) is calculated as follows:

in is the normalized category weight, which reflects the scarcity of the category. The fewer the category samples, the higher the weight. v(i) is the interaction value of item i. α is an adjustment parameter used to control the impact of the item interaction value on the final weight.

6. The method according to claim 4, characterized in that step S62 specifically comprises:

A long-tail oriented contrastive learning optimization strategy is designed. On the basis of obtaining the category balance weight W _opt (i), a dynamic sampling cooling mechanism is used for subset sampling. Then, cross-layer contrastive learning is used to encourage the model to achieve more fine-grained hierarchical information alignment. The optimized joint loss function is as follows:

For a given user-item pair, we not only focus on the representation z _i at the final layer, but also on the embeddings of the intermediate layers By incorporating contrastive learning into the contrastive objectives and implementing contrastive learning between multiple levels of embedded representations, the model embedding in the representation space can capture more three-dimensional features; when finally calculating the loss, the main task (BPR) and contrastive learning (CL) are combined for joint optimization. The main task loss focuses on optimizing the user's preference ranking of positive and negative samples to improve the recommendation accuracy; the contrastive learning loss aims to enhance the diversity and robustness of the representation space and improve the ability to capture long-tail items; through this joint training method, the model can not only maintain good performance in terms of recommendation accuracy, but also effectively improve diversity and robustness.

7. An intelligent recommendation device for collaborative diversity and accuracy, characterized in that it includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the intelligent recommendation method for collaborative diversity and accuracy described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that a program is stored thereon, and when the program is executed by a processor, the intelligent recommendation method of collaborative diversity and accuracy described in any one of claims 1-6 is implemented.