Cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination

Abstract

The invention provides a cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination, which designs a variational self-supervised coding module to optimize edge user and item feature representations, and adopts a maximized mutual information strategy to distinguish different edge user and item feature representations. The problem of insufficient edge user information is effectively relieved, so that the data sparsity difference between the source domain and the target domain is reduced, and the negative migration problem is relieved; secondly, for a user partial overlapping recommendation scene, a dynamic pseudo-overlapping judgment mechanism is provided in a dynamic pseudo-overlapping judgment module, two-domain overlapping users and non-overlapping users are clustered to obtain a pseudo-overlapping user set, the area of the overlapping users can be maximized, a source domain and a target domain are effectively connected, and therefore the cross-domain recommendation performance is improved, and finally, the recommendation efficiency is improved. In the feature reconstruction and recommendation module, more accurate prediction interaction scores of a source domain and a target domain can be output, and finally cross-domain recommendation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a cross-domain recommendation method based on variational self-supervisory coding and dynamic pseudo-overlap determination. Background Art

[0002] Cross-Domain Recommendation (CDR) methods use the rich interactive information in the source domain as an aid to improve the data sparsity problem in the target domain. Deep learning technologies represented by Graph Neural Network (GNN) are widely used in the field of cross-domain recommendation. However, due to the sparsity difference between the source and target domain data, blindly migrating the feature preferences of the source domain to the target domain sometimes not only fails to improve the recommendation performance of the target domain, but also reduces the accuracy of the recommendation, resulting in a negative transfer problem, which makes the rich information in the source domain unable to be directly used in the target domain.

[0003] However, the traditional approach to alleviating the negative transfer problem cannot effectively alleviate the negative transfer problem because it does not consider the existence of marginal users as the root cause of the sparsity differences between different domains. At the same time, the existing CDR methods generally assume that the source domain users and the target domain users completely overlap in the recommendation scenario, which is not universal. Non-overlapping user information is often difficult to effectively extract in cross-domain tasks, affecting the cross-domain recommendation performance. Summary of the invention

[0004] To this end, an embodiment of the present invention proposes a cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination to solve the problem that the prior art cannot effectively alleviate negative transfer and affects the cross-domain recommendation performance.

[0005] A cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlapping determination according to an embodiment of the present invention is applied to a cross-domain recommendation system, wherein the cross-domain recommendation system includes a variational self-supervised coding module, a dynamic pseudo-overlapping determination module, and a feature reconstruction and recommendation module; The method comprises: In the variational self-supervised encoding module, based on the content features of users and projects, user temporary neighborhood representations and project temporary neighborhood representations are generated, and then the user and project potential feature matrices in the source domain, target domain and shared domain are constructed through the user temporary neighborhood representations and project temporary neighborhood representations. Based on the user and project potential feature matrices in the source domain, target domain and shared domain, an inference network is introduced in the variational reasoning process to solve the vector representation of edge users based on the posterior distribution. Based on the vector representation of edge users based on the posterior distribution, the collaborative features between edge users and non-edge users are learned. In the generation process, the potential variables of edge users are combined with the latent variables of edge users through the MLP network. JaccardMap the content attributes of non-marginal users whose matrix similarity is greater than the threshold, generate the content features reconstructed by the marginal users, and finally maximize the collaborative features and the content features reconstructed by the marginal users to obtain the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix; In the dynamic pseudo-overlapping determination module, the fuzzy C-means clustering algorithm is used to group overlapping users to obtain overlapping user groups. The leader influence method is introduced into the overlapping user groups, and the Pearson correlation calculation formula and the trust calculation formula are combined to calculate the influence of the leader in the overlapping user group. The similarity and correlation between the leader and the non-overlapping users in the source domain, and the similarity and correlation between the leader and the non-overlapping users in the target domain are solved respectively. Finally, the upper and lower limits of the correlation are obtained by using the attribute weak concept similarity to limit the number of pseudo-overlapping users, thereby obtaining the pseudo-overlapping user set. In the feature reconstruction and recommendation module, the influence weights between leaders and users, and between users in the pseudo-overlapping user set are calculated, and the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix are combined to obtain the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation. Based on the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation, the predicted interaction score of the source domain and the predicted interaction score of the target domain are output, and project recommendation lists are provided to users in the source domain and the target domain respectively according to the predicted interaction score of the source domain and the predicted interaction score of the target domain.

[0006] According to the cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlapping judgment of an embodiment of the present invention, a variational self-supervised coding module is designed to optimize the feature representation of edge users and projects, and a maximum mutual information strategy is adopted to distinguish different feature representations of edge users and projects, effectively alleviating the problem of insufficient information of edge users, thereby reducing the data sparsity difference between the source domain and the target domain, and alleviating the negative transfer problem; secondly, for the user partial overlapping recommendation scenario, a dynamic pseudo-overlapping judgment mechanism is proposed in the dynamic pseudo-overlapping judgment module, and the overlapping users and non-overlapping users of the two domains are clustered to obtain a pseudo-overlapping user set, which can maximize the overlapping user area and effectively connect the source domain and the target domain, thereby improving the cross-domain recommendation performance; finally, in the feature reconstruction and recommendation module, a more accurate source domain and target domain user node embedding representation, as well as the source domain and target domain project node embedding are obtained, so that more accurate source domain and target domain predicted interaction scores can be output, and finally cross-domain recommendation is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The above and / or additional aspects and advantages of the embodiments of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 is a schematic diagram of the structure of a cross-domain recommendation system according to an embodiment of the present invention; Figure 2 It is a flowchart of a cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to an embodiment of the present invention. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0009] The embodiment of the present invention proposes a cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap judgment, which is applied to a cross-domain recommendation system. Figure 1 The cross-domain recommendation system includes a variational self-supervised encoding module, a dynamic pseudo-overlapping determination module, and a feature reconstruction and recommendation module.

[0010] See also Figure 2 , the method comprises steps S1 to S3: Step S1, in the variational self-supervised encoding module, based on the content features of users and projects, generate user temporary neighborhood representations and project temporary neighborhood representations, then construct user and project potential feature matrices in the source domain, target domain and shared domain through the user temporary neighborhood representations and project temporary neighborhood representations, based on the user and project potential feature matrices in the source domain, target domain and shared domain, introduce an inference network in the variational reasoning process to solve the vector representation of edge users based on the posterior distribution, learn the collaborative features between edge users and non-edge users based on the vector representation of edge users based on the posterior distribution, and in the generation process, compare the potential variables of edge users with the latent variables of edge users through the MLP network. Jaccard The content attributes of non-marginal users whose matrix similarity is greater than the threshold are mapped to generate content features reconstructed by marginal users. Finally, the collaborative features and the content features reconstructed by marginal users are maximized to obtain the source domain user latent feature matrix, the source domain project latent feature matrix, the target domain user latent feature matrix, and the target domain project latent feature matrix.

[0011] Unlike homogeneous graphs, each node in a bipartite graph is of a different type from its neighboring nodes. Therefore, it is not appropriate to directly update the node embedding by aggregating the features of its one-hop neighbors. To this end, the present invention learns each node embedding from its two-hop neighbors at each layer. Through the mean operation (MEAN) with nonlinear transformation, the first k The temporary user and item neighborhood representation is as follows: ; ; in, It is k Layer User u b The temporary neighborhood representation of It is k Layer Items i a The temporary neighborhood representation of is the LeakyReLU activation function, It is k Layer user weight matrix, It is k Layer item weight matrix, MEAN represents the mean operation, Is a user Content characteristics, It is a project Content characteristics, It is k -1st level users Content characteristics, It is k -1st floor project Content characteristics, Is a user u b The even-hop neighbor set of It is a project i a The set of even-hop neighbors of .

[0012] Through the generated temporary user and item neighborhood representations, domain-specific node embeddings can be generated. Taking the user representation of the source domain as an example, in the process of constructing the user and item latent feature matrices in the source domain, target domain, and shared domain through the user temporary neighborhood representation and the item temporary neighborhood representation, the following formula is satisfied: ; ; in, is the mean vector of source domain users, is the standard deviation vector of source domain users, and All of them conform to the Gaussian distribution. Norm is the row normalization function, is the transposed matrix, is the concatenation operator, If it is a source domain user, is the Softplus function, , , , is the parameter matrix.

[0013] pass and The potential feature matrix of source domain users Take samples; The user u b As a source domain user, combined with the variational Bayes reparameterization technique, we have: ; in, Is a user u b The latent feature vector of Is a user u b The mean vector of Is a user u b The standard deviation vector of is the element-wise product, is a normal Gaussian noise vector.

[0014] Through the above process, the information propagation of homogeneous neighbors of users in the source domain is preliminarily realized, and the potential feature matrix of users in the source domain is obtained. Based on the same principle, the potential feature matrix of the source domain project can be obtained , target domain user potential feature matrix , target domain item latent feature matrix .

[0015] In the bipartite heterogeneous graph, only user / item nodes with even-number-hop neighbors are considered, and the influence of edge user features on preferences is not paid attention to, which has certain limitations. Therefore, the present invention alleviates the problem of insufficient information of edge users through variational self-supervision. First, a collaborative feature representation of edge users is established, and the matrix similarity ratio between them and non-edge users is solved based on the collaborative feature representation; then, a variational autoencoder (VAE) is used to generate a mapping between the user's own collaborative features and other user content features with high similarity ratios; finally, the collaborative features and inherent content features generated between the mutual information are maximized to distinguish different edge user representations.

[0016] The variational autoencoder includes a variational reasoning part and a generation part. In the variational reasoning process, an inference network is introduced to solve the vector representation of edge users based on the posterior distribution, satisfying the following formula: ; ; ; in, Is a marginal user Based on the vector representation of the posterior distribution, Is a marginal user All attribute codes of is the standard deviation operation, is the inference network, Is a marginal user The mean vector of Is a marginal user The standard deviation vector of is a diagonal function, yes and of Jaccard Matrix similarity, when When it is greater than the threshold, and There are collaborative features between them.

[0017] Marginal users and non-marginal users For example, , of Jaccard When the matrix similarity is greater than the set threshold, it is considered and There is a collaborative feature relationship. Non-marginal users with high similarity The content feature of a is represented as the concatenation of all its attribute encodings.

[0018] In the generation process, the latent variables of edge users are combined with Jaccard The content attribute mapping of non-marginal users whose matrix similarity is greater than the threshold generates marginal users Reconstructed content features .

[0019] The goal of variational inference is to optimize the free variational parameters to minimize the KL divergence. The present invention uses the user latent variables obtained by the mean square error to sample and reparameterize, and the loss function of variational inference is As shown in the formula: ; in, yes KL Divergence, represents the prior latent variable distribution, Is a user All attribute codes of E Express expectations.

[0020] Mutual information measures the correlation between random variables. When encoding content features into the latent space, in order to improve the ability to distinguish content features, the maximum mutual information ( MIM ) strategy further enhances the correlation between the potential content features and their corresponding collaborative features. The present invention maximizes the loss function of the mutual information strategy Designed for: ; in, yes sigmoid function, is a vector input function.

[0021] The feature self-supervision task is assisted by the variational inference loss and the mutual information maximization loss. Therefore, by maximizing the collaborative features and the content features reconstructed by the edge users, the source domain user potential feature matrix is ​​obtained. , source domain project potential feature matrix , target domain user potential feature matrix , target domain item latent feature matrix In the process, the following equation is satisfied: ; in, is the overall loss function of the variational self-supervised encoding module.

[0022] Step S2, in the dynamic pseudo-overlapping determination module, the overlapping users are grouped using the fuzzy C-means clustering algorithm to obtain overlapping user groups, the leader influence method is introduced into the overlapping user groups, and the Pearson correlation calculation formula and the trust calculation formula are combined to calculate the influence of the leader in the overlapping user group, and the similarity and correlation between the leader and the non-overlapping users of the source domain, and the similarity and correlation between the leader and the non-overlapping users of the target domain are solved respectively, and finally the upper and lower limits of the correlation are obtained using the attribute weak concept similarity, and the number of pseudo-overlapping users is limited, thereby obtaining a pseudo-overlapping user set.

[0023] Among them, let the overlapping user set be O , divide O into multiple groups, and then define the group centers. In each iteration, the fuzzy C-means (FCM) algorithm calculates a membership matrix and a set of cluster centers, and then solves the objective function value based on them to group overlapping users and obtain overlapping user groups.

[0024] The Pearson correlation coefficient (PCC) is widely used in similarity measurement, and users' evaluations of common items are usually used as a measure of their similarity.

[0025] Specifically, the Pearson correlation calculation formula is: ; in, Indicates user and users The similarity between For users About Project Ratings, For users About Project Rating, For users The average rating of For users The average rating of is a set of items, if is positive, then the user and users Similar, otherwise not similar.

[0026] To calculate the trust value between users, the present invention uses an implicit asymmetric trust measurement method to make up for the preference distance problem ignored by PCC. It uses a linear combination method to calculate the intersection and distance of the rating items between users. Specifically, the trust calculation formula is: ; ; ; in, Indicates user and users The degree of trust between Is a user and users The asymmetric trust between Is a user and users The rating distance between Indicates user Itemsets with non-zero scores, Indicates user Itemsets with non-zero scores, Indicates user and users The set of items that have no score equal to zero.

[0027] After obtaining the overlapping user groups, it is necessary to determine the leader in the group. The present invention introduces a leader influence method to select the user with the highest member trust and similarity as the leader of the group and guide the opinions of the members to point to common interests.

[0028] Specifically, the expression for calculating the influence of a leader in an overlapping user group is: ; in, Represents leader Influence within overlapping user groups, Represents leader and users The similarity between Represents leader and users The degree of trust between Indicates the number of users in the overlapping user group.

[0029] When the leader in the group is obtained, the higher the correlation, the more likely it is to become a pseudo-overlapping user. The similarity and correlation between the leader and the non-overlapping users in the source domain, and the similarity and correlation between the leader and the non-overlapping users in the target domain are solved. If the correlation is higher than the set value, it can be defined as a pseudo-overlapping user. The calculation formula of similarity and similarity is: Satisfy the following formula: ; ; in, Represents leader With non-overlapping users The similarity of Represents leader With non-overlapping users The correlation degree, Indicates non-overlapping users Itemsets with non-zero scores, Represents leader Itemsets with non-zero scores, For non-overlapping users About Project Rating, For non-overlapping users The average rating of For leaders About Project Rating, Indicates non-overlapping users The set of even-hop neighbors of items with non-zero scores, Represents leader The set of even-hop neighbors of itemsets with non-zero scores.

[0030] From the above content, we can get the correlation between specific domain users and overlapping users, but we still need to draw a boundary. Here, we choose the attribute weak concept similarity to solve the upper and lower limits of the correlation, so as to limit the number of pseudo-overlapping users.

[0031] Specifically, the upper and lower limits of the correlation are obtained by using the attribute weak concept similarity, and the following formula is satisfied in the process of limiting the number of pseudo-overlapping users: ; ; in, Indicates the upper limit similarity of the attribute weak concept, Indicates the lower limit similarity of the attribute weak concept.

[0032] At this point, the overlapping user area is expanded through the dynamic pseudo-overlapping determination mechanism, and pseudo-overlapping users can be used to effectively connect non-overlapping users in two domains.

[0033] Step S3, in the feature reconstruction and recommendation module, calculate the influence weights between leaders and users, and between users in the pseudo-overlapping user set, and combine the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix to obtain the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation; output the predicted interaction score of the source domain and the predicted interaction score of the target domain based on the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation; and provide project recommendation lists to users in the source domain and the target domain respectively according to the predicted interaction score of the source domain and the predicted interaction score of the target domain.

[0034] The influence of the leader on the members is not constant, but varies according to each member's trust and similarity to the leader. In addition, the influence weight of each member is also variable, and the influence weights between members are asymmetric.

[0035] Specifically, in the process of calculating the influence weights between leaders and users, and between users in the pseudo-overlapping user set, the following formula is satisfied: ; ; in, Represents leader For users The influence weight of Indicates user For users The influence weight of For users With leaders The correlation degree, Represents leader and users The degree of trust between Represents leader and users The similarity between After judging the pseudo-overlapping users and the dual influence weights of the two domains, the present invention can obtain more accurate embedding representations of user nodes in the source domain and the target domain to alleviate the cold start problem of some users caused by data sparsity.

[0036] Specifically, in the process of obtaining the source domain user node embedding representation and the target domain user node embedding representation, the following formula is satisfied: ; ; in, is the source domain user node embedding representation, is the target domain user node embedding representation, is the linear transformation matrix, is the leader potential characteristics matrix, is the leadership influence weight, Is a user in the source domain The latent feature matrix, Is a user in the target domain The latent feature matrix.

[0037] In order to reduce the error between the user's latest predicted preference and the user's actual preference and improve the accuracy of user features and model performance, the present invention uses a modified cross entropy loss function to optimize the model and uses softmax The function outputs the predicted item score.

[0038] Specifically, in the process of outputting the predicted interaction score of the source domain and the predicted interaction score of the target domain based on the source domain user node embedding representation, the target domain user node embedding representation, the source domain item node embedding representation, and the target domain item node embedding representation, the following formula is satisfied: ; ; in, is the predicted interaction score of the source domain, is the predicted interaction score of the target domain, is the source domain item node embedding representation, is the target domain item node embedding representation, is the normalized exponential function, Yes The node embedding representation obtained after transposition processing is Yes The node embedding representation obtained after transposition.

[0039] In this embodiment, the total loss function of the feature reconstruction and recommendation module is for: ; ; ; in, is the source domain loss function, is the target domain loss function, are network parameters, , , is the weight coefficient, is the set of observed interactions, is a certain number of negative instances randomly sampled from the unobserved interaction set to prevent overfitting; are the interaction scores of users and items in the known source domain, is the interaction score between users and items in the known target domain. and To control the balance of the sample, use Control sample regularization, network parameters Used to prevent overfitting.

[0040] According to the above-mentioned cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlapping judgment, a variational self-supervised coding module is designed to optimize the feature representation of edge users and projects, and the maximum mutual information strategy is adopted to distinguish different feature representations of edge users and projects, which effectively alleviates the problem of insufficient information of edge users, thereby reducing the data sparsity difference between the source domain and the target domain and alleviating the negative transfer problem; secondly, for the user partial overlapping recommendation scenario, a dynamic pseudo-overlapping judgment mechanism is proposed in the dynamic pseudo-overlapping judgment module, and the overlapping users and non-overlapping users of the two domains are clustered to obtain the pseudo-overlapping user set, which can maximize the overlapping user area and effectively connect the source domain and the target domain, thereby improving the cross-domain recommendation performance. Finally, in the feature reconstruction and recommendation module, more accurate source domain and target domain user node embedding representations, as well as source domain and target domain project node embeddings are obtained, which can output more accurate predicted interaction scores of the source domain and target domain, and finally realize cross-domain recommendation.

[0041] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0042] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination, characterized in that: Applied to a cross-domain recommendation system, the cross-domain recommendation system includes a variational self-supervised encoding module, a dynamic pseudo-overlapping determination module, and a feature reconstruction and recommendation module; The method comprises: In the variational self-supervised encoding module, based on the content features of users and projects, user temporary neighborhood representations and project temporary neighborhood representations are generated, and then the user and project potential feature matrices in the source domain, target domain and shared domain are constructed through the user temporary neighborhood representations and project temporary neighborhood representations. Based on the user and project potential feature matrices in the source domain, target domain and shared domain, an inference network is introduced in the variational reasoning process to solve the vector representation of edge users based on the posterior distribution. Based on the vector representation of edge users based on the posterior distribution, the collaborative features between edge users and non-edge users are learned. In the generation process, the potential variables of edge users are combined with the latent variables of edge users through the MLP network. Jaccard Map the content attributes of non-marginal users whose matrix similarity is greater than the threshold, generate the content features reconstructed by the marginal users, and finally maximize the collaborative features and the content features reconstructed by the marginal users to obtain the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix; In the dynamic pseudo-overlapping determination module, the fuzzy C-means clustering algorithm is used to group overlapping users to obtain overlapping user groups. The leader influence method is introduced into the overlapping user groups, and the Pearson correlation calculation formula and the trust calculation formula are combined to calculate the influence of the leader in the overlapping user group. The similarity and correlation between the leader and the non-overlapping users in the source domain, and the similarity and correlation between the leader and the non-overlapping users in the target domain are solved respectively. Finally, the upper and lower limits of the correlation are obtained by using the attribute weak concept similarity to limit the number of pseudo-overlapping users, thereby obtaining the pseudo-overlapping user set. In the feature reconstruction and recommendation module, the influence weights between leaders and users, and between users in the pseudo-overlapping user set are calculated, and the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix are combined to obtain the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation. Based on the source domain user node embedding representation, the target domain user node embedding representation, the source domain project node embedding representation, and the target domain project node embedding representation, the predicted interaction score of the source domain and the predicted interaction score of the target domain are output, and project recommendation lists are provided to users in the source domain and the target domain respectively according to the predicted interaction score of the source domain and the predicted interaction score of the target domain.

2. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 1 is characterized in that: In the variational self-supervised encoding module, in the process of generating user temporary neighborhood representation and item temporary neighborhood representation based on the content features of the user and the item, the following equation is satisfied: ； ； in, It is k Layer User u b The temporary neighborhood representation of It is k Layer Items i a The temporary neighborhood representation of is the LeakyReLU activation function, It is k Layer user weight matrix, It is k Layer item weight matrix, MEAN represents the mean operation, Is a user content features, It is a project content features, It is k -1st level users content features, It is k -1st floor project content features, Is a user u b The even-hop neighbor set of It is a project i a The set of even-hop neighbors of ; In the process of constructing the user and item potential feature matrices in the source domain, target domain, and shared domain through the user temporary neighborhood representation and the item temporary neighborhood representation, the following formula is satisfied: ； ； in, is the mean vector of source domain users, is the standard deviation vector of source domain users, Norm is the row normalization function, is the transposed matrix, is the concatenation operator, If it is a source domain user, is the Softplus function, , , , is the parameter matrix; pass and The potential feature matrix of source domain users Take samples; The user u b As a source domain user, you have: ； in, Is a user u b The latent feature vector of Is a user u b The mean vector of Is a user u b The standard deviation vector of is the element-wise product, is a normal Gaussian noise vector.

3. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 2 is characterized in that: In the process of variational inference, the inference network is introduced to solve the vector representation of edge users based on the posterior distribution, satisfying the following formula: ； ； ； in, Is a marginal user Based on the vector representation of the posterior distribution, Is a marginal user All attribute codes of is the standard deviation operation, is the inference network, Is a marginal user The mean vector of Is a marginal user The standard deviation vector of is a diagonal function, yes and of Jaccard Matrix similarity, when When it is greater than the threshold, and There are collaborative features between them; In the generation process, the latent variables of edge users are combined with Jaccard The content attribute mapping of non-marginal users whose matrix similarity is greater than the threshold generates marginal users Reconstructed content features .

4. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 3 is characterized in that: In the process of maximizing the collaborative features and the content features reconstructed by the edge users, obtaining the source domain user potential feature matrix, the source domain project potential feature matrix, the target domain user potential feature matrix, and the target domain project potential feature matrix, the following formula is satisfied: ； ； ； in, is the overall loss function of the variational self-supervised encoding module, is the loss function for variational inference, is the loss function of the strategy to maximize mutual information, yes KL Divergence, represents the prior latent variable distribution, Is a user All attribute codes of E Express expectations, yes sigmoid function, is a vector input function.

5. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 3 is characterized in that: The Pearson correlation calculation formula is: ； in, Indicates user and users The similarity between For users About Project Rating, For users About Project Rating, For users The average rating of For users The average rating of It is a set of items; The trust calculation formula is: ； ； ； in, Indicates user and users The degree of trust between Is a user and users The asymmetric trust between Is a user and users The rating distance between Indicates user Itemsets with non-zero scores, Indicates user Itemsets with non-zero scores, Indicates user and users The set of items whose scores are all non-zero; The expression for calculating the influence of a leader within overlapping user groups is: ； in, Represents leader Influence within overlapping user groups, Represents leader and users The similarity between Represents leader and users The degree of trust between Indicates the number of users in the overlapping user group; In the process of solving the similarity and correlation between the leader and the non-overlapping users in the source domain, and the similarity and correlation between the leader and the non-overlapping users in the target domain, the following formula is satisfied: ； ； in, Represents leader With non-overlapping users The similarity of Represents leader With non-overlapping users The correlation degree, Indicates non-overlapping users Itemsets with non-zero scores, Represents leader Itemsets with non-zero scores, For non-overlapping users About Project Rating, For non-overlapping users The average rating of For leaders About Project Rating, Indicates non-overlapping users The set of even-hop neighbors of items with non-zero scores, Represents leader The set of even-hop neighbors of itemsets with non-zero scores.

6. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 5, characterized in that: Using the weak concept similarity of attributes to obtain the upper and lower limits of the correlation, and in the process of limiting the number of pseudo-overlapping users, the following formula is satisfied: ； ； in, Indicates the upper limit similarity of the attribute weak concept, Indicates the lower limit similarity of the attribute weak concept.

7. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 6, characterized in that: In the process of calculating the influence weights between leaders and users, and between users in the pseudo-overlapping user set, the following formula is satisfied: ； ； in, Represents leader For users The influence weight of Indicates user For users The influence weight of For users With leaders The correlation degree, Represents leader and users The degree of trust between Represents leader and users The similarity between In the process of obtaining the source domain user node embedding representation and the target domain user node embedding representation, the following formula is satisfied: ； ； in, is the source domain user node embedding representation, is the target domain user node embedding representation, is the linear transformation matrix, is the leader potential characteristics matrix, is the leadership influence weight, Is a user in the source domain The latent feature matrix, Is a user in the target domain The latent feature matrix of In the process of outputting the predicted interaction score of the source domain and the predicted interaction score of the target domain based on the source domain user node embedding representation, the target domain user node embedding representation, the source domain item node embedding representation, and the target domain item node embedding representation, the following formula is satisfied: ； ； in, is the predicted interaction score of the source domain, is the predicted interaction score of the target domain, is the source domain item node embedding representation, is the target domain item node embedding representation, is the normalized exponential function, Yes The node embedding representation obtained after transposition processing is Yes The node embedding representation obtained after transposition.

8. The cross-domain recommendation method based on variational self-supervised coding and dynamic pseudo-overlap determination according to claim 7, characterized in that: The total loss function of the feature reconstruction and recommendation module for: ； ； ； in, is the source domain loss function, is the target domain loss function, are network parameters, , , is the weight coefficient, is the set of observed interactions, is a certain number of negative instances randomly sampled from the set of unobserved interactions, are the interaction scores of users and items in the known source domain, are the interaction scores of users and items in the known target domain.

Images