An adversarial generative bidirectional cross-domain recommendation method in a cold start recommendation scenario
By employing an adversarial generative bidirectional cross-domain recommendation method, which utilizes generative adversarial networks and relational graph convolutional networks to capture personalized features of users and items, the method addresses the performance limitations of recommendations in cold start scenarios and achieves more efficient recommendation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANSHAN UNIV
- Filing Date
- 2024-07-18
- Publication Date
- 2026-07-31
AI Technical Summary
Existing cross-domain recommendation methods suffer from unstable mapping functions and negative transfer issues in cold start scenarios, and fail to effectively capture user preferences within domains and differences between domains, resulting in limited recommendation performance.
We adopt an adversarial generative bidirectional cross-domain recommendation method, which uses generative adversarial networks to generate user preference representations, and combines relational graph convolutional networks and attention mechanisms to capture personalized features of users and items. We improve recommendation performance through multi-task optimization.
It enhances the scalability and generalization ability of the recommendation model, improves the recommendation performance for cold-start users, captures the fine-grained differences in user preferences, and alleviates the negative transfer problem.
Smart Images
Figure CN118861424B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of recommender system technology, and in particular to an adversarial generative bidirectional cross-domain recommender method for cold-start recommendation scenarios. Background Technology
[0002] With the development of the internet and the surge in digital content, users are facing an explosive increase in the amount of information they encounter. In this context, it becomes difficult for users to find content or products of interest from this vast amount of information. Recommendation systems have emerged to help users efficiently filter and select information. By mining and analyzing user behavioral data (such as browsing history, click records, purchase records, search keywords, etc.), recommendation systems build user interest models and recommend personalized content or products to users based on these models. This personalized filtering mechanism can greatly improve the efficiency with which users find content of interest.
[0003] Numerous studies have found that traditional recommender systems still face two major challenges: data sparsity and the cold start problem. Cross-domain recommendation leverages the rich source domain information of interactive items to improve recommendation performance for users with cold starts in the target domain, providing a new solution to alleviate these problems. Among existing cross-domain recommendation methods, a mainstream approach is based on mapping functions, exemplified by EMCDR. This method primarily learns a mapping function to transfer overlapping user preferences from the source domain to the target domain. Considering the different preferences of different users, PTUPCDR (Personalized Transfer of User Preferences for Cross-domain Recommendation) introduces meta-learning concepts to extend mapping function-based methods. It uses a meta-network to train a different personalized mapping function for each user, capturing personalized preferences.
[0004] However, the mapping functions learned by meta-learning methods are often unstable and difficult to adapt to new tasks. Furthermore, most existing cross-domain models rely solely on source domain information to assist recommendation tasks in the target domain. Considering the differences between domains, unidirectional cross-domain methods often lead to negative transfer problems. Many studies have shown that both the source and target domains contain valuable information, and adopting a dual-domain mutual reinforcement strategy can help alleviate the negative transfer problem.
[0005] Existing dual-domain cross-domain recommendation methods do not take into account users' preferences for different items within a domain or the inter-domain differences when users migrate preferences between domains. The representational ability of dual-domain user preference modeling is insufficient, resulting in limited recommendation performance in both the source and target domains. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide an adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios, which utilizes generative adversarial networks to generate powerful user preference representations and enhance recommendation performance in both domains.
[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0008] An adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios includes the following steps:
[0009] S1. Preprocess the comment data to obtain user-project interaction sequence information;
[0010] S2. Pre-train the user-item interaction graph based on the user-item interaction sequence;
[0011] S3. Utilize pre-trained embeddings to perform personalized interaction feature representations for dual-domain users;
[0012] S4. Utilize personalized interaction features to generate a dual-domain hybrid user preference representation;
[0013] S5. Adversarial learning is applied to the dual-domain hybrid user preference representation.
[0014] S6. Use a multi-task joint optimization approach to train the model and obtain the best user representation and item representation in both domains.
[0015] S7. Using the finally obtained user and item representations, personalized recommendations are made to users in both domains.
[0016] A further improvement to the technical solution of this invention lies in: In S2, the different ratings given by users to different items and the different ratings assigned to items by different users are captured by a relational graph convolutional network to capture user representations and item representations of the source and target domains.
[0017] S2 includes the following steps:
[0018] S21. Represent the source / target domain users and items using vectors;
[0019] Considering that users rate different items differently, a relational graph convolutional network is used to capture the complex relationships in the heterogeneous graph, resulting in representative user and item representations.
[0020]
[0021] in, and Let N be the feature representations of node u and node v at layer l. r (u) and N r(v) represents the items that interact with the current user under the rating relationship r, and the users who rate the items under the rating relationship r, respectively. and All of these are learnable parameters during the training process, and σ is the activation function between layers. Here, tanh is used as the activation function.
[0022] S22, Multi-level Vector Aggregation:
[0023]
[0024]
[0025] In this context, it is assumed that the relational graph convolutional network has L layers, || represents the connection operation between layers, and the user representation e is obtained after the connection operation through multiple layers. u and project representation e v ;
[0026] S23. Optimization of pre-trained vector representation;
[0027] Define the matrix factorization loss function:
[0028]
[0029] Where |R| represents the number of user-item ratings, r uv This represents the rating of user u for item v. After optimization of the matrix factorization task, the best user embedding and item embedding are output to prepare for personalized feature transfer.
[0030] S24. Obtain the pre-trained source domain user embeddings. Source domain project embedding Target domain user embedding Target domain project embedding
[0031] A further improvement to the technical solution of the present invention is that S3 includes the following steps:
[0032] S31. Introduce a simple and effective attention mechanism to calculate the interaction weight matrix between source domain users and source domain items, and the interaction weight matrix between target domain users and target domain items, respectively:
[0033]
[0034] in, W represents the user embedding vector from the source / target domain obtained in S2. att Here, b is the weight parameter, and b is the bias parameter. For source / target domain projects v d For source / target domain user u dPersonalized transferable weights, where d∈{s,r};
[0035] S32. Using the weight matrix obtained in S31, calculate the personalized transferable feature embeddings of users in the source and target domains using a weighted approach:
[0036]
[0037] in, This represents the attention weights of users' preferences for different items in the source / target domain. For each item that the current user interacts with, the weighted average is used to obtain the user's final personalized transferable feature embedding.
[0038] A further improvement to the technical solution of the present invention is that S4 includes the following steps:
[0039] S41. Target domain sampling: Randomly sample users with multiple interactions in the target domain as guide users for cold-start knowledge transfer. This process is described as follows:
[0040] U k ={u∈U t |n u ≥k}
[0041] Among them, U k U represents the target domain guided user set obtained from sampling. t n represents the set of all users in the target domain. u This represents the number of interactions user u has in the target domain, and k represents the sampling threshold.
[0042] S42. Learn the transfer weight matrix for each domain separately. This process is described as follows:
[0043]
[0044] in, For mixed source / target domain weights, Personalized transferable feature embeddings of source / target domains, u i u k All are guiding users for target domain sampling, W p With W k All are learnable parameters;
[0045] S43. Generation of hybrid user preference representations;
[0046] By aggregating the user's mixed source domain weights, the final user mixed preference representation is calculated. The aggregation operation is represented as follows:
[0047]
[0048] Among them, U k Indicates from guiding user set Uk A subset of guided users sampled in the middle, W q For learnable parameters, For representing user interests in hybrid source / target domains;
[0049] S44. Hybrid Source / Target Domain User Preference Representation Optimization:
[0050]
[0051] S45. Introduce cosine loss function for training optimization to capture the similarity of user preferences between the two domains;
[0052] The optimization process is as follows:
[0053]
[0054] Where d(·) is the cosine distance.
[0055] A further improvement to the technical solution of the present invention is that S5 includes the following steps:
[0056] S51. Introduce a domain discriminator to capture the differences in interests among users in both domains. The domain discriminator is defined as follows:
[0057]
[0058] S52. To obtain the final embedding of user interests in the target domain, the domain discriminator optimization process is as follows:
[0059]
[0060] Where, d l Represents the domain label, when d l When d = 0, it represents the source domain; when d = 0, it represents the source domain. l =1 indicates the target domain; D(·) represents the domain discriminator defined in S51.
[0061] A further improvement to the technical solution of this invention lies in: In S6, the optimization of the recommendation task and the personalized representation adversarial generation task, the multi-task joint optimization function is defined as follows:
[0062] L rec =λ1L s_rec +λ2L t_rec
[0063] L adv =L cos +L dis
[0064] L=μ(L rec )+(1-μ)(L adv ).
[0065] A further improvement to the technical solution of the present invention is that S7 includes the following steps:
[0066] S71. Calculate the dot product between the final user representation and the item representation in the source domain and the target domain, respectively, and use it as the prediction score.
[0067]
[0068] S72. Sort the predicted scores for each domain in descending order, select the top 10 products, and recommend them to relevant users in the source / target domain.
[0069] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:
[0070] 1. The present invention designs an adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios. This method addresses the negative transfer problem in cross-domain recommendation systems by changing the method from using only source domain knowledge to enhance the target domain to simultaneously using knowledge from both domains to enhance each other. This improves the recommendation performance of both domains and enhances the scalability and generalization ability of the recommendation model.
[0071] 2. The present invention designs an adversarial generative bidirectional cross-domain recommendation method for cold start recommendation scenarios. In order to capture users' personalized preferences within a single domain, it uses a relation-aware graph neural network to aggregate item representations with different ratings into user representations, thereby enhancing the representation ability of user preferences.
[0072] 3. The present invention designs an adversarial generative bidirectional cross-domain recommendation method for cold start recommendation scenarios. By introducing a generative adversarial learning paradigm, it captures the fine-grained differences in user preferences between domains and alleviates the negative transfer problem. Attached Figure Description
[0073] Figure 1 This is a schematic diagram of the model flow in this invention;
[0074] Figure 2 This is a model framework diagram in this invention. Detailed Implementation
[0075] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:
[0076] like Figure 1 , 2 As shown, an adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios includes the following steps:
[0077] S1. Preprocess the comment data to obtain user-project interaction sequence information;
[0078] Using Amazon's public review dataset, we selected datasets from three domains: music, books, and movies. We analyzed the data features of the three domains and constructed three different cross-domain recommendation tasks: music to books, music to movies, and movies to books. Based on the different recommendation tasks, we obtained the corresponding overlapping users and user-item interaction sequence information.
[0079] S2. Pre-train the user-item interaction graph based on the user-item interaction sequence;
[0080] This study uses a relational graph convolutional network to capture user and item representations in the source and target domains, reflecting how users rate different items and how different users assign different ratings to items. This prepares for subsequent transferable personalized feature learning;
[0081] S2 includes the following steps:
[0082] S21. Represent the source / target domain users and items using vectors;
[0083] Considering that users have different ratings for different items, a relational graph convolutional network is used to capture the complex relationships in the heterogeneous graph and obtain representative user and item representations.
[0084]
[0085] in, and Let N be the feature representations of node u and node v at layer l. r (u) and N r (v) represents the items that interact with the current user under the rating relationship r, and the users who rate the items under the rating relationship r, respectively. and All of these are learnable parameters during the training process, and σ is the activation function between layers. Here, tanh is used as the activation function.
[0086] S22, Multi-level Vector Aggregation:
[0087]
[0088] In this context, it is assumed that the relational graph convolutional network has L layers, and || represents the concatenation operation between layers. After the concatenation operation, the user representation e after multiple layers of connections is obtained. u and project representation e v ;
[0089] S23. Optimization of pre-trained vector representation;
[0090] Define the matrix factorization loss function:
[0091]
[0092] Where |R| represents the number of user-item ratings, r uv This represents the rating of user u for item v. After optimization of the matrix factorization task, the best user embedding and item embedding are output to prepare for personalized feature transfer.
[0093] S24. Obtain the pre-trained source domain user embeddings. Source domain project embedding Target domain user embedding Target domain project embedding
[0094] S3. Utilize pre-trained embeddings to perform personalized interaction feature representations for dual-domain users;
[0095] Considering that different user interactions with items contribute differently to user preferences, a simple and effective attention mechanism is introduced, which uses user embeddings and item embeddings from the source and target domains to learn personalized transferable features for the source and target domains, respectively.
[0096] S3 includes the following steps:
[0097] S31. Introduce a simple and effective attention mechanism to calculate the interaction weight matrix between source domain users and source domain items, and the interaction weight matrix between target domain users and target domain items, respectively:
[0098]
[0099] in, W represents the user embedding vector from the source / target domain obtained in S2. att Here, b is the weight parameter, and b is the bias parameter. For source / target domain projects v d For source / target domain user u d Personalized transferable weights, where d∈{s,r};
[0100] S32. Using the weight matrix obtained in S31, calculate the personalized transferable feature embeddings of users in the source and target domains using a weighted approach:
[0101]
[0102] in, This represents the attention weights of users' preferences for different items in the source / target domain. For each item that the current user interacts with, the weighted average is used to obtain the user's final personalized transferable feature embedding.
[0103] S4. Utilize personalized interaction features to generate a dual-domain hybrid user preference representation;
[0104] To enhance the knowledge transfer capabilities of cold-start users, hybrid source domain representations and hybrid target domain representations are learned separately to generate indistinguishable dual-domain user hybrid representations. This facilitates the transfer of source domain user preferences to target domain users and prepares for subsequent training of the domain discriminator.
[0105] S4 includes the following steps:
[0106] S41. Target domain sampling: Randomly sample users with multiple interactions in the target domain as guide users for cold-start knowledge transfer. This process is described as follows:
[0107] U k ={u∈U t |n u ≥k}
[0108] Among them, U k U represents the target domain guided user set obtained from sampling. t n represents the set of all users in the target domain. u This represents the number of interactions user u has in the target domain, and k represents the sampling threshold.
[0109] S42. Learn the transfer weight matrix for each domain separately. This process is described as follows:
[0110]
[0111] in, For mixed source / target domain weights, Personalized transferable feature embeddings of source / target domains, u i u k All are guiding users for target domain sampling, W p With W k All are learnable parameters;
[0112] S43. Generation of hybrid user preference representations;
[0113] By aggregating the user's mixed source domain weights, the final user mixed preference representation is calculated. The aggregation operation is represented as follows:
[0114]
[0115] Among them, U k Indicates from guiding user set U k A subset of guided users sampled in the middle, W q For learnable parameters, For representing user interests in hybrid source / target domains;
[0116] S44. Hybrid Source / Target Domain User Preference Representation Optimization:
[0117]
[0118] S45. Introduce cosine loss function for training optimization to capture the similarity of user preferences between the two domains;
[0119] The optimization process is as follows:
[0120]
[0121] Where d(·) is the cosine distance.
[0122] S5. Adversarial learning is applied to the dual-domain hybrid user preference representation.
[0123] To distinguish the interest representations of users in different domains and capture the fine-grained differences between different domains, this method adopts a generative adversarial learning paradigm. It uses a domain discriminator to determine whether the generated personalized hybrid user representation can correctly represent the user's true preferences. High-quality user interest representations are generated in a game-like manner for target domain recommendation prediction.
[0124] S5 includes the following steps:
[0125] S51. Introduce a domain discriminator to capture the differences in interests among users in both domains. The domain discriminator is defined as follows:
[0126]
[0127] S52. To obtain the final embedding of user interests in the target domain, the domain discriminator optimization process is as follows:
[0128]
[0129] Where, d l Represents the domain label, when d l When d = 0, it represents the source domain; when d = 0, it represents the source domain. l =1 indicates the target domain; D(·) represents the domain discriminator defined in S51.
[0130] S6. Use a multi-task joint optimization approach to train the model and obtain the best user representation and item representation in both domains.
[0131] S6 includes the following steps:
[0132] The joint optimization function for optimizing recommendation tasks and personalized representation adversarial generation tasks is defined as follows:
[0133] L rec =λ1L s_rec +λ2L t_rec
[0134] L adv =L cos +Ldis
[0135] L=μ(L rec )+(1-μ)(L adv )
[0136] S7. Using the finally obtained user and item representations, personalized recommendations are made to users in both domains.
[0137] S7 includes the following steps:
[0138] S71. Calculate the dot product between the final user representation and the item representation in the source domain and the target domain, respectively, and use it as the prediction score.
[0139]
[0140] S72. Sort the predicted scores for each domain in descending order, select the top 10 products, and recommend them to relevant users in the source / target domain.
[0141] In summary, this invention addresses the problems of existing dual-domain cross-domain recommendation methods failing to consider users' preferences for different items within a domain and the inter-domain differences in user preference migration between domains, resulting in insufficient representational power in dual-domain user preference modeling and thus limited recommendation performance in both the source and target domains. Furthermore, it utilizes generative adversarial networks to generate powerful user preference representations, thereby enhancing recommendation performance within both domains.
Claims
1. An adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios, characterized in that: Includes the following steps: S1. Preprocess the comment data to obtain user-project interaction sequence information; S2. Pre-train the user-item interaction graph based on the user-item interaction sequence; S3. Utilize pre-trained embeddings to perform personalized interaction feature representations for dual-domain users; S4. Utilize personalized interaction features to generate a dual-domain hybrid user preference representation, specifically including the following steps: S41. Target domain sampling: Randomly sample users with multiple interactions in the target domain as guiding users for cold-start knowledge transfer, described as follows: in, This indicates the target domain obtained from sampling, guiding the user set. Represents the set of all users in the target domain. Indicates user Number of interactions in the target domain Indicates the sampling threshold; S42. Learn the transfer weight matrix for each domain, described as follows: in, For mixed source / target domain weights, Embedding personalized transferable features for the source / target domain. , All of these are guiding users who sample the target domain. and All are learnable parameters; S43. Generation of hybrid user preference representations; By aggregating the user's mixed source domain weights, the final user mixed preference representation is calculated. The aggregation operation is represented as follows: in, 'Indicates from the user group being guided A subset of guided users sampled in the middle. For learnable parameters, For representing user interests in hybrid source / target domains; S44. Hybrid Source / Target Domain User Preference Representation Optimization: S45. Introduce cosine loss function for training optimization to capture the similarity of user preferences between the two domains; The optimization process is as follows: in, Cosine distance; S5. Adversarial learning is performed on dual-domain hybrid user preference representations; S6. Use a multi-task joint optimization approach to train the model and obtain the best user representation and item representation in both domains. S7. Using the finally obtained user and item representations, personalized recommendations are made to users in both domains.
2. The adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios according to claim 1, characterized in that: In S2, user ratings for different items and the different ratings assigned to items by different users are captured using a relational graph convolutional network to obtain user and item representations in the source and target domains. ); S2 includes the following steps: S21. Represent the source / target domain users and items using vectors; Considering that users rate different items differently, a relational graph convolutional network is used to capture the complex relationships in the heterogeneous graph, resulting in representative user and item representations. in, and They are nodes and nodes In the Layer feature representation, and These respectively represent the rating relationship The following items interact with the current user and their rating relationships Below are the users who rated the project. and These are all learnable parameters during the training process. This is the activation function between layers; here, tanh is used as the activation function. S22, Multi-level Vector Aggregation: Among them, it is assumed that the relational graph convolutional network has a total of layer, This represents the connection operations between layers, resulting in a multi-layered user representation. and project representation ; S23. Optimization of pre-trained vector representation; Define the matrix factorization loss function: in, This indicates the number of user-project ratings. Indicates user For the project The rating, after optimization of the matrix factorization task, outputs the best user embedding and item embedding, preparing for personalized feature transfer; S24. Obtain the pre-trained source domain user embeddings. Source domain project embedding Target domain user embedding Target domain project embedding .
3. The adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios according to claim 1, characterized in that: S3 includes the following steps: S31. Introduce a simple and effective attention mechanism to calculate the interaction weight matrix between source domain users and source domain items, and the interaction weight matrix between target domain users and target domain items, respectively: in, This represents the user embedding vector obtained from the source / target domain in S2. For weight parameters, This is the bias parameter; S32. Using the weight matrix obtained in S31, calculate the personalized transferable feature embeddings of users in the source and target domains using a weighted approach: in, This represents the attention weights of users' preferences for different items in the source / target domain. For each item that the current user interacts with, the weighted average is used to obtain the user's final personalized transferable feature embedding.
4. The adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios according to claim 1, characterized in that: S5 includes the following steps: S51. Introduce a domain discriminator to capture the interest differences between dual-domain users. The domain discriminator is defined as follows: S52. To obtain the final embedding of user interests in the target domain, the domain discriminator optimization process is as follows: in, Indicates a field label, when When, it represents the source domain; when Time represents the target domain; This refers to the domain discriminator defined in S51.
5. The adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios according to claim 1, characterized in that: In S6, the joint optimization function for the recommendation task and the personalized representation adversarial generation task is defined as follows: 。 6. The adversarial generative bidirectional cross-domain recommendation method for cold-start recommendation scenarios according to claim 1, characterized in that: S7 includes the following steps: S71. Calculate the dot product between the final user representation and the item representation in the source domain and the target domain, respectively, and use it as the prediction score. S72. Sort the predicted scores for each domain in descending order, select the top 10 items, and recommend them to relevant users in the source / target domain.