Multi-view soft contrast learning method for personalized recommendation

Abstract

The invention discloses a multi-view soft contrast learning method for personalized recommendation, and belongs to the field of link prediction, graph neural networks and contrast learning tasks. A user-project interaction view, a user-user social contact view and a project-project interaction view are utilized, and node embedding is constructed through a view encoder. A weighted aggregation information spreading scheme is adopted in the views, cross iteration operation is carried out between the views, high-order connectivity between the views is introduced, and embedding comprehensiveness and accuracy are improved. And a transfer learning method is introduced to learn the personalized preference of the user. And finally, the soft label is integrated into a comparison loss function, so that the model can learn from strong association and weak association, and potential information is captured. Experimental results show that the prediction performance and generalization ability of the recommendation system can be improved, more accurate and comprehensive support is provided for applications such as the recommendation system and intelligent search, and more accurate and personalized recommendation is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of link prediction, graph neural network and contrastive learning tasks, and in particular to a multi-view soft contrastive learning method for personalized recommendation. Background Art

[0002] The rapid expansion of Internet data resources has made the problem of information overload more prominent. In order to enable users to find interesting content faster, build a bridge between information production and consumption, attract attention, efficiently match content or products that users like, and enhance user experience and stickiness, recommendation systems are widely used on network platforms. Recommendation systems can recommend content or services that users may be interested in based on information such as past behavior and interest tendencies. Because recommendation systems help companies accurately predict user needs and make scientific decisions, they can provide users with more convenient personalized services. At present, they have been applied in many fields such as music, movies, short videos, news, and e-commerce. For example, Amazon, as an e-commerce website, uses recommendation systems to recommend products that users may be interested in. Its recommendation system generates recommendation results based on information such as user purchases, searches, and browsing history. In this case, improving the performance of recommendation systems has become a hot research topic.

[0003] Recommendation algorithm is the core part of recommendation system. Traditional recommendation algorithms include content-based recommendation, collaborative filtering-based recommendation and hybrid recommendation. Such methods usually only rely on the historical behavior of users, such as the user's click record, purchase record and rating record of items, to generate recommendation results. However, in the actual application of recommendation system, the items that actually interact with users only account for a small number in the recommendation system, which easily leads to the problem of data sparsity. Due to the lack of sufficient data support, it is difficult for the algorithm to effectively recommend long-tail items that are only concerned by a few users. In addition, when new users or items join the system, due to the lack of available historical interaction data, the recommendation algorithm cannot know the user's interests and preferences, and it is difficult to make effective recommendations, which in turn affects the performance indicators of the recommendation system. In addition, due to the relatively simple methods and models of traditional recommendation algorithms, it is difficult to explain the decision-making process and the basis of recommendation results, and there is a problem of poor interpretability. This will make users distrust the recommendation results, thereby affecting the application effect of the recommendation algorithm and user experience.

[0004] In view of the above limitations and challenges, this paper proposes a multi-view soft contrastive learning method for personalized recommendation. Our method simultaneously utilizes the user-item interaction view, the user-user social view, and the item-item relationship view to construct user embeddings and item embeddings through a heterogeneous multi-view embedding method. A weighted aggregation information propagation scheme is adopted within the view, and a cross-iterative information propagation scheme is adopted between views, which introduces high-order connectivity between views, thereby improving the comprehensiveness and accuracy of the embedding. A transfer learning method is introduced to learn the personalized preferences of users and items. Finally, we integrate soft labels into the contrastive loss function, allowing the model to learn from strong and weak associations and capture the underlying information. This allows for more accurate prediction performance. Summary of the invention

[0005] The technical problem solved by the present invention is: in view of the problems of insufficient accuracy in obtaining information from multiple views, fusion and alignment of information in multiple views, and recommendation tasks, a multi-view soft contrast learning method for personalized recommendation is proposed, so as to fully learn the topological details in multiple views, mine the high-order connectivity and correlation between views, fully understand and capture the complex semantics and contextual information in the views, provide embedding accuracy and comprehensiveness, thereby improving the prediction performance and generalization ability of the recommendation system, providing more accurate and comprehensive support for applications such as recommendation systems and intelligent search, and achieving more accurate and personalized recommendations.

[0006] The technical solution of the present invention is: a multi-view soft contrast learning method for personalized recommendation, the steps are as follows: (1) Perform global embedding initialization on the user nodes and item nodes in the user-item interaction view to obtain the original embedding. (2) The original embedding is input into different view encoders to obtain node embeddings on different views. (3) Intra-view message passing is performed on three different views based on a designed attention message propagation method that does not require additional transformation. (4) Perform cross-iterative message propagation operations among the three views. (5) Use a meta-network to aggregate embeddings from different stages across three views. (6) Use bridge functions to capture the direct and indirect effects between users and items. (7) Fusion is performed based on the embeddings obtained at each stage to obtain a more accurate node embedding (8) Learning and training are performed according to the designed multi-view soft contrast learning module to obtain the final node embedding. (9) Generate the final prediction results.

[0007] Furthermore, the user nodes and project nodes in the user-project interaction view are initialized with global embedding to obtain the original embedding. For structured knowledge such as user and project embedding, the global initialization embedding method is used.

[0008] Furthermore, user embedding and item embedding construct more flexible node embeddings in different views based on different view encoders, further revealing the differences in interaction patterns between user-item interaction view, user-user social view and item-item relationship view.

[0009] Furthermore, a lightweight GCN using simple aggregation operations is designed to capture the relationship between nodes, and on this basis, an attention mechanism is added to give different importance to different neighboring nodes, thereby obtaining an attention message propagation method that does not require additional conversion. This method is used for aggregation in a bottom-up manner, and the neighbor information of each layer is regarded as input into the network in hierarchical order to obtain the sequential information in the view, and the attention mechanism is used to illustrate the importance of these neighbor information to generate node embedding.

[0010] Furthermore, due to the different information levels of different views, different node embedding outputs can be obtained. These node embeddings play different roles in different views. In order to enhance the overall semantic representation, cross-iteration operations are used to make these direct or indirect relationships reflected in the node embeddings.

[0011] Furthermore, since the above steps generate node embeddings in different views, we utilize a meta-network to aggregate the initial embeddings, neighbor information, and view embeddings to enhance the stability and expressiveness of the model and ensure a comprehensive understanding of node features and relationships.

[0012] Furthermore, the bridge function is used to customize the learning of user social interaction, personal preferences, and the impact of connections between items on the embedding of specific users or items.

[0013] Furthermore, to enhance the influence of user-item interactions on node embeddings, we utilize a weighted sum to fuse the above node embeddings, which means that the final node embedding acknowledges that users’ preferences are formed by their interaction history and social network.

[0014] Furthermore, an effective multi-view contrastive learning paradigm is proposed, which considers the similarities between positive sample pairs under different views and assigns the probability or confidence of potential links to negative samples, which means that the model can learn and capture the underlying information from strong and weak associations and fuse them into node embeddings.

[0015] Recommended modeling is performed on the embeddings between nodes to generate the final prediction results.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention simultaneously utilizes the user-item interaction view, the user-user social view and the item-item relationship view, and obtains the user embedding and the item embedding by constructing a multi-view embedding method. A weighted aggregation information propagation scheme is adopted within the view range, and a cross-iteration information propagation scheme is adopted between views, which enhances the high-order connectivity and correlation between views, thereby improving the comprehensiveness and accuracy of the embedding. A transfer learning method is introduced to describe the characteristics of users and items. Finally, we integrate the soft labels into the contrastive loss function, allowing the model to learn and capture the underlying information from strong and weak associations, thereby obtaining the final node embedding. Finally, a multi-layer perceptron MLP is used to make the final recommendation judgment on the node embedding to obtain the most ideal prediction valuation. In this process, we fully learn the topological details and hierarchical information of the three views, explore the valuable characteristics of the heterogeneous topological structures in the three views, better understand and capture the complex semantics and contextual information within the view, and improve the robustness and adaptability through soft allocation and soft boundary loss functions. Experimental results show that the present invention can effectively alleviate the problems of data sparsity, insufficient utilization of heterogeneous topological information, and unsatisfactory accuracy of recommendation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to make the technical solution of the present invention clear, some drawings will be selected below to illustrate the present invention. The following drawings are only some examples of the present invention; users can obtain other similar drawings based on the drawings. Among them: Figure 1 This is a flow chart of a multi-view soft contrast learning method for personalized recommendation according to the present invention; Figure 2 It is the overall module diagram of the present invention; Figure 3 To implement the environment diagram; Figure 4 Calculate the result graph for the model; DETAILED DESCRIPTION

[0018] The following will be a more in-depth and complete description of the technical solutions in the implementation cases of the present invention. The implementation cases are part of the implementation cases. Users can obtain other implementation cases based on the present invention without creative work. The "other implementation cases" all belong to the scope of protection of the present invention.

[0019] like Figure 1 As shown, a multi-view soft contrast learning method for personalized recommendation includes the following steps: (1) Embedding initialization of nodes in the user-item interaction view. (2) Construct node embeddings in the other two views based on the view encoder. (3) Message passing within the view. (4) Message exchange between views. (5) Meta-network aggregation messages. (6) Customize learning using bridge functions. (7) Node embedding fusion. (8) Use soft contrastive learning to train the model and obtain the final node embedding. (9) Generate the final prediction results.

[0020] According to one embodiment of the present invention, combined with the present invention Figure 2 The framework provides a simple module explanation of the present invention. A multi-view soft contrast learning method for personalized recommendation consists of three main modules: a heterogeneous graph representation learning module, a meta-knowledge transfer learning module, and a soft contrast learning module. The first module is the heterogeneous graph representation learning module, whose main function is to capture the representation learning of nodes through self-gated units, intra-view weighted aggregation, and cross-iteration between views, revealing the differences in interaction patterns between views. The second module is the meta-knowledge transfer learning module, which is mainly responsible for promoting personalized dependency modeling across multiple views. The third module is the soft contrast learning module, which integrates soft labels into the soft contrast loss and considers the similarity between positive and negative sample pairs under different views to enhance the learning effect of the model.

[0021] Based on the above implementation, the method in the above step (1) is: embedding and initializing the users and items appearing in the user-item interaction view. In order to obtain the global structural information of the heterogeneous graph, the initial representation of each node should reflect its characteristics in the entire graph. We use the most common global initialization to represent users and items in a continuous vector space. as well as

[0022] The method in step (2) above is: In order to further reveal the differences in interaction patterns between views, the original embedding is filtered and weighted using a pre-filter with a self-gating unit. We input the original embedding into different view encoders, taking the user embedding in the user-user social view as an example, the specific formula is: in, represents the initial embedding of the user in the user-user social view, σ(·) represents the sigmoid nonlinear function, ⊙ represents the element-wise multiplication, and W uu and b uu represents the learnable transformation and bias parameters. The item embeddings in the item-item relational view are also obtained in a similar way.

[0023] The method of message passing within the view in step (3) above is to use a lightweight GCN with a simple aggregation operation, which can also effectively capture node relationships. On this basis, we design an attention message propagation paradigm that does not require additional transformations or nonlinear activation functions. We take the user-item interaction view as an example for detailed modeling, and the operations in other views are the same process. The calculation formula for message propagation within the user-item interaction view is as follows: Where A is the weight assigned to the user, a u ∈A, is a learnable query, is a learnable key. represents the l-th layer input embedding of the user node in the user-item interaction view, where Used as a scaling factor to normalize the dot product and reduce the risk of vanishing gradients during training. N represents the output embedding of item node i in the l-th iteration. i and N u Represents the neighbor nodes of user u and item i respectively. Input elements in the view It is also obtained through a similar process.

[0024] In the above step (4), the method of message exchange between views is as follows: in order to obtain the heterogeneous semantics of high-order connections from multi-hop links and capture the influence flow in the three views, the message propagation is cross-iterated between the three views. The output embedding of the lth layer of the user-item interaction view is used as the input embedding of the l+1th layer of the user-user social view and the project-item relationship view. The output embedding of the l+1th layer of the user-user social view and the project-item relationship view is used as the input embedding of the l+2th layer of the user-item interaction view. The specific formula is expressed as: Embed is the output of the lth layer of the messaging module from the user-user social view, and represents the input of the l+1th layer in the user-item interaction view, and The definition of is similar.

[0025] The way the meta-network aggregates messages in step (5) above is that users’ decisions are influenced by their preferences, social relationships, and item associations. Our model aims to adapt to these influences and exploit subtle user behaviors. To achieve this, we use a meta-network to transfer knowledge between various data views, enhancing our ability to learn from auxiliary information sources. The implementation details are as follows: where || represents the concatenation operation that merges multiple view representations. The concatenated embedding H uRepresents the user’s multi-view meta-knowledge. u The user-item interaction view reflects the user’s intrinsic preferences. Since the user-user social view encapsulates the social relationships in the user network, E uu Represents the influence of the user's social network. is additional neighborhood information that explicitly enhances the modeling of direct graph connections. i The construction of is similar to this one. By connecting insights from different viewpoints, we enrich the embedding learning process and deepen our understanding of the complex interactions between users and items under various behavioral paradigms.

[0026] The method of using bridge function to customize learning in step (6) above is: Since bridge function helps to encapsulate and transmit the subtle behavioral preferences shown by a single user, we build a bridge function for user preferences. The specific formula is: g 1 (·) and g 2 (·) consists of two fully connected layers using the PReLU activation function. The matrix and is derived from user meta-knowledge H u The derived custom transformation matrix is ​​designed to build a bridge function for user preferences. δ(·) is the PReLU activation function. Through these parameterized bridge functions, Not only the direct interactions between users and items are captured, but also potential indirect effects, such as the influence of friends of friends on user preferences.

[0027] In the above step (7), the method of node embedding fusion is: In order to enhance the effect of user-item interaction, we use weighted sum. Specifically, it is: represents the end-user embedding, r u represents the influence of user-item interactions. This means that the final embedding acknowledges that users’ preferences are formed by their interaction history and social network, applying the same process to

[0028] In the above step (8), the soft contrast learning training model is used to obtain the final node embedding method: the multi-view soft contrast learning loss based on infonce is used to enhance the representation learning of users and items. We first take the user's presentation loss as an example and define the user's positive sample loss. The specific formula is: u' represents a negative sample. s(·) is the cosine similarity function. τ represents the temperature coefficient, which can automatically identify difficult negative samples. Similarly, we can get the information loss of the positive sample, which is recorded as We set the confidence score of the positive sample pair to 1, and the confidence score of the negative sample pair is determined by the distance between the nodes, as follows: The value range of α is limited to [0,1], which represents the upper bound for distinguishing negative sample pairs that are close to each other. Here, ε is a hyperparameter that controls the clarity, and D(·) represents the distance calculation function, and we use the cosine distance. Therefore, the loss of negative sample pairs is: The total soft contrastive learning loss is: akin, It is also obtained in the same way.

[0029] The method for generating the final prediction result in the above step (9) is: recommend modeling the final path embedding between nodes to generate the final prediction result

[0030] In summary, this paper proposes a multi-view soft contrastive learning method for personalized recommendation, which aims to capture the interactive heterogeneity and similarity between nodes in such networks. In order to improve the performance of the recommendation system, we enhance the comprehensiveness and accuracy of embeddings through weighted aggregation within views, iteration across multiple views, and soft contrastive learning. In this process, the topological details in multiple views are fully learned, the high-order connectivity and correlation between views are mined, the complex semantics and contextual information in the views are fully understood and captured, and the accuracy and comprehensiveness of the embedding are provided, thereby improving the predictive performance and generalization ability of the recommendation system, providing more accurate and comprehensive support for applications such as recommendation systems and intelligent search, and achieving more accurate and personalized recommendations.

[0031] The above is only one embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-view soft contrast learning method for personalized recommendation, characterized in that: The task method comprises the following steps: (1) Perform global embedding initialization on the user nodes and item nodes in the user-item interaction view to obtain the original embedding. (2) The original embedding is input into different view encoders to obtain node embeddings on different views. (3) Based on a designed attention message propagation method that does not require additional transformation, intra-view message passing is performed simultaneously on three different views. (4) Perform cross-iterative message propagation operations among the three views. (5) Use a meta-network to aggregate embeddings from different stages across three views. (6) Use bridge functions to capture the direct and indirect effects between users and items. (7) Fusion is performed based on the embeddings obtained at each stage to obtain a more accurate node embedding (8) Learning and training are performed according to the designed multi-view soft contrast learning module to obtain the final node embedding. (9) Generate the final prediction results.

2. The multi-view soft contrast learning method for personalized recommendation according to claim 1, characterized in that: In step (2), node embeddings of different views are obtained through different view encoders. Specifically, the original user embeddings obtained in the user-item interaction view are converted into and the original project embedded Input into different view encoders respectively, for example, input into the view encoder of user-user social view, following the formula: Get user embedding in user-user social view Where W uu and b uu Represents the learnable transformation and bias parameters. And so on, we can get the item embedding in the item-item relationship view 3. The multi-view soft contrast learning method for personalized recommendation according to claim 1, characterized in that: In step (3), according to a designed attention message propagation method that does not require additional conversion, message transmission within the view is performed simultaneously on three different views. Specifically, we designed a lightweight GCN that uses a simple aggregation operation to capture node relationships, and on this basis, added an attention mechanism that assigns different importance to different adjacent nodes, thereby obtaining an attention message propagation method that does not require additional conversion. Taking the user-item interaction view as an example for detailed modeling, the specific formula is: Where A is the weight assigned to the user, a u ∈A, is a learnable query, is a learnable key. represents the l-th layer input embedding of the user node in the user-item interaction view, where Normalizing the dot product as a scaling factor can reduce the risk of vanishing gradients during training. N represents the output embedding of item node i in the l-th iteration. i and N u Represents the neighbor nodes of user u and item i respectively. Input elements in the view It is also obtained through a similar process.

4. The multi-view soft contrast learning method for personalized recommendation according to claim 1, characterized in that: In step (4), a cross-iterative message propagation operation is performed among the three views. Specifically, the output embedding of the user-item interaction view at layer l is used as the input embedding of the user-user social view and the item-item relationship view at layer l+1. The output embedding of the user-user social view and the item-item relationship view at layer l+1 is used as the input embedding of the user-item interaction view at layer l+2.

5. The multi-view soft contrast learning method for personalized recommendation according to claim 1, characterized in that: In step (8), learning and training are performed according to the designed multi-view soft contrast learning module to obtain the final node embedding, which specifically includes: considering the similarity between positive sample pairs under different views, and assigning the probability or confidence of potential links to negative samples to enhance the learning effect of the model.

Images