A Cross-Domain Recommendation Method Based on Adaptive Deentanglement

Abstract

This invention discloses a cross-domain recommendation method based on adaptive disentanglement, belonging to the field of recommendation system technology, including the following steps: S1, data acquisition; S2, constructing a user-item interaction graph; S3, initializing user and item representations; S4, using variational bidirectional graph encoding (VBGE) to encode the interaction graphs of the source domain, target domain, and shared domain respectively, and using VBGE for information propagation to generate latent variable representations of users and items, capturing knowledge between the source and target domains; S5, employing adaptive disentanglement technology to obtain the shared information dimension of users in the source domain; S6, employing shared information enhancement technology to improve the recommendation effect of the target and source domains; S7, personalized recommendation; This invention effectively separates shared information and domain-specific information through adaptive disentanglement technology to improve the accuracy and transferability of recommendations and avoid negative knowledge transfer.

Description

Technical Field

[0001] This invention belongs to the field of recommendation system technology, specifically relating to a cross-domain recommendation method based on adaptive deentanglement. Background Technology

[0002] With the rapid development of the internet and the explosive growth of information, users face an unprecedented problem of information overload. In this era of information overload, effectively filtering content that users are interested in from massive amounts of data has become a huge challenge. Recommendation systems, as an important means of solving the problem of information overload, have been widely used in e-commerce, video platforms, social media, and other fields. By analyzing user behavior data and item information, recommendation systems can provide users with personalized content recommendations, helping them efficiently obtain the information they need.

[0003] Traditional recommender systems face several key challenges, particularly data sparsity and the cold start problem. Data sparsity refers to the fact that most entries in the user-item interaction matrix are missing, making it difficult to accurately infer user interests from limited interaction data. The cold start problem refers to the difficulty for recommender systems to provide accurate recommendations when new users or items have not yet accumulated sufficient interaction data. These two problems significantly limit the effectiveness of traditional recommender systems.

[0004] Collaborative filtering fills data gaps by analyzing the similarity of user behavior. It assumes that if two users have exhibited similar preferences and behavioral patterns in the past, they will likely have similar interests in the future. In this way, collaborative filtering algorithms can use the group of users most similar to the target user's behavior to predict the target user's preferences, thereby improving recommendation accuracy. However, collaborative filtering-based recommendation methods still suffer from data sparsity and cold-start problems, especially in scenarios lacking sufficient user behavior data, where the effectiveness of the recommendation system is difficult to guarantee.

[0005] Cross-domain recommendation methods, on the other hand, improve recommendation performance by integrating data and knowledge from multiple domains and leveraging the correlations between different domains. Specifically, cross-domain recommendation compensates for the information shortage caused by data sparsity by sharing user behavior data or item attribute information across domains, and provides more reliable recommendation solutions by utilizing existing knowledge in the source domain when facing a cold start in a new domain. For example, user preferences in the book domain can provide a reference for movie recommendations, thereby effectively alleviating the problems of data sparsity and cold start.

[0006] However, existing cross-domain recommendation methods still have many limitations. One key problem is that information in different domains is often entangled with shared information and domain-specific information. Many existing models fail to adequately distinguish between certain shared features that are beneficial to multiple domains and certain specific features that only apply to a single domain when performing knowledge transfer between domains. For example, the book and movie domains may share the feature of "genre," but "writing style" in books may not be effectively transferred in the movie domain. This mixture of domain-specific information can easily lead to negative transfer, weakening the recommendation effect.

[0007] Therefore, there is a need for a cross-domain recommendation method based on adaptive disentanglement that can finely handle the separation of domain-specific information and shared information, improve the accuracy and transferability of recommendations. Summary of the Invention

[0008] The purpose of this invention is to provide a cross-domain recommendation method based on adaptive deentanglement, which can finely handle the separation of domain-specific information and shared information, and effectively improve the accuracy and transferability of recommendations.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A cross-domain recommendation method based on adaptive disentanglement includes the following steps:

[0011] S1. Data Acquisition: Acquire historical interaction data of the source domain and target domain, analyze and select overlapping users of the source domain and target domain, and extract the corresponding user-item interaction data from the source domain and target domain respectively based on the above overlapping users.

[0012] S2. Construct User-Item Interaction Graph: Using the overlapping user and item interaction data of the source and target domains, construct user-item interaction graphs for the source, target, and shared domains. The user-item interaction graph is a heterogeneous bipartite graph, which includes user type nodes and item type nodes. The interaction graphs of the source and target domains use users and items from their respective domains, while the interaction graph of the shared domain includes users and items from both the source and target domains.

[0013] S3. User and item representation initialization: User and item representations are initialized separately for the source and target domains, respectively. For the shared domain, only the user representation is initialized, and its item representation is combined with the item representations of the source and target domains.

[0014] S4. Variational Bidirectional Graph Coding (VBGE): VBGE is used to encode the interaction graphs of the source domain, target domain, and shared domain respectively. Information is propagated through VBGE to generate approximate posterior distributions of users and items. Latent variable representations of users and items are generated by summarizing homogeneous information, capturing knowledge between the source and target domains.

[0015] S5. Adaptive deentanglement: Adaptive deentanglement technique is used to separate the shared information and domain-specific information in the source domain user representation and the same user representation in the shared domain to obtain the shared information dimension of the source domain user.

[0016] S6. Shared Information Enhancement: The shared information enhancement technique is used to enhance the user representation of the target domain by de-entanglement of the source domain and the shared domain, and at the same time, the user representation of the source domain is enhanced by de-entanglement of the target domain and the shared domain, thereby improving the recommendation effect of the target domain and the source domain.

[0017] S7. Personalized Recommendation: The final user representation and item representation are multiplied by a dot product to determine the user's interest in the item, and personalized recommendations are provided to the user based on this interest.

[0018] A further improvement of the technical solution of the present invention is that: in step S2, for the interaction graph of the source domain and the target domain, the users are overlapping users and the items are items within their respective domains; for the interaction graph of the shared domain, the users are overlapping users and the items consist of items from the source domain and the target domain.

[0019] A further improvement to the technical solution of this invention lies in the following: In step S4, VBGE encodes the user-project interaction graph through a two-step information propagation mechanism, including the following steps:

[0020] S401, Homogeneous Information Propagation: Since the user-project interaction graph is a heterogeneous bipartite graph, VBGE adopts a two-step information propagation mechanism to propagate information between homogeneous neighbors in order to generate intermediate representations;

[0021] S402, Intermediate Representation Calculation: For the user representation of the source domain, calculate its intermediate representation based on 2-hop neighbors.

[0022]

[0023] Where Norm(·) is the row normalization operation, A X Let δ(·) be the source domain user-project interaction matrix, and let δ(·) be the Leaky ReLU activation function. The learnable parameter matrix;

[0024] S403, Latent Variable Distribution Modeling: Based on Intermediate Representation Calculate the distribution of latent variables for users:

[0025]

[0026] in, For the concatenation operation, φ(·) is the Softplus function. and For learnable parameter matrix, and These are the mean and standard deviation represented by the user, respectively;

[0027] S404, Latent Variable Sampling: Sampling the user's latent variable representation using the reparameterization trick.

[0028]

[0029] in, A source domain-specific representation for the user, used to capture the user's characteristics in that domain.

[0030] A further improvement to the technical solution of the present invention is that: in step S5, calculations are performed using formulas (5) and (6), and the dimension of the shared information is determined by a threshold:

[0031]

[0032] in, For the user's representation distribution, D KL To obtain the KL divergence operation, δ i Let τ be the KL divergence across various dimensions between users, and δ be the threshold. i <τ, determine the dimension and number of dimensions where the user-shared information is located.

[0033] A further improvement of the technical solution of the present invention is as follows: In step S6, the average value of the representation distribution of users in the source domain or target domain and the corresponding representation distribution in the shared domain is first calculated based on the determined shareable dimension to obtain the average user representation distribution. Then, the user representation distribution is resampled to obtain the specific representation of shareable user information. Finally, the user information is combined with the corresponding user representation to enhance the information contained in the user representation.

[0034] A further improvement of the technical solution of the present invention is that: in step S6, the process of enhancing shared information is described as shown in formulas (7) and (8):

[0035]

[0036] user_share=mask⊙z i (8)

[0037] Where μ is the mean of the distribution, σ is the standard deviation of the distribution, and ∈ i To obtain from the standard normal distribution The noise term in the mid-sample, z i This represents the result of sampling the distribution; mask indicates the shareable dimension, and user_share is the shareable information obtained after de-entanglement.

[0038] A further improvement of the technical solution of the present invention is as follows: In step S7, the shared representation and the domain-specific representation de-entangled by the user are first fused by weighted summation. After the user representation and the item representation are generated, the user's interest in the item is measured by dot product calculation. Then, the interest is normalized and nonlinear transformation is performed by sigmoid activation function to ensure that the calculated interest is suitable for ranking recommendation. Finally, all candidate items are ranked according to the normalized interest, and the top N items with the highest interest are selected as the recommendation results.

[0039] A further improvement to the technical solution of the present invention is that step S7 includes the following steps:

[0040] S701. Generation of User Representations: In cross-domain recommendation scenarios, user representations consist of shared representations and domain-specific representations, which are fused using a weighted summation method to fully utilize cross-domain information while maintaining domain personalization. The calculation formula for user representations is as follows:

[0041]

[0042] Among them, E u This represents the final user representation vector; This represents the decoupled user-shared representation, which represents users' common interest information across all domains; This represents the user's original personalized representation in the current target domain, reflecting the user's specific preferences in that domain;

[0043] S702. Calculate user interest in the project: In user representation E u After generation, the corresponding project representation is obtained, and the dot product is used to measure the user's interest in the project:

[0044] E i =E[i] (10)

[0045] s(u,i)=E u ·E i (11)

[0046] in, The item representation matrix obtained from model training, where N is the total number of items, d is the embedding dimension, and i is the item index, Ei Let be the representation vector of item i; s(u,i) represents the degree of interest of user u in item i;

[0047] S703. Normalization: The interest scores are normalized to ensure that the calculated interest scores s(u,i) are suitable for ranking and recommendation. A sigmoid activation function is used for non-linear transformation.

[0048]

[0049] The interest level should be within the range of (0,1);

[0050] S704, Personalized Recommendation: Based on normalized interest levels All candidate items are sorted, and the top N items with the highest interest are selected as the recommendation results, i.e.:

[0051]

[0052] Then, the top N ranked projects are recommended to the user to achieve personalized recommendations.

[0053] A further improvement of the technical solution of the present invention is that, in step S1, overlapping users refer to users who have appeared in both the source domain and the target domain.

[0054] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0055] This invention is based on an adaptive deentanglement cross-domain recommendation method. By using adaptive deentanglement technology, it effectively separates shared information and domain-specific information to improve the accuracy and transferability of recommendations and avoid negative knowledge transfer.

[0056] This invention employs adaptive deentanglement technology to effectively separate shared information from domain-specific information in different domains, quantifies and identifies shared features between the source and target domains, improves the accuracy of cross-domain recommendations, and thus effectively enhances the performance of recommendation systems in the face of data sparsity and cold start problems, optimizes recommendation performance, and demonstrates significant advantages in multiple practical application scenarios.

[0057] This invention can accurately model users' cross-domain interests and preferences, optimize personalized recommendation effects, and improve user experience; by mining shared information, it can effectively alleviate the problem of data sparsity in a single domain and improve the applicability of the recommendation system in the case of insufficient data.

[0058] This invention can also improve the cold start problem. Even if a user has no historical behavior in the target domain, it can provide high-quality recommendations by leveraging information from other domains, enhancing the recommendation effect for new users and new projects. At the same time, this invention supports cross-platform recommendations, such as joint recommendations between social media and e-commerce platforms, thereby improving user conversion rates.

[0059] This invention performs information separation and fusion directly in the embedded space. Compared with traditional cross-domain transfer learning methods, it has higher computational efficiency, reduces computational costs in large-scale applications, and makes it more suitable for practical scenarios such as e-commerce, online education, and short video recommendation, thus having broad commercial application value. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the model flow in the cross-domain recommendation method based on adaptive disentanglement of the present invention;

[0061] Figure 2 This is a model framework diagram of the cross-domain recommendation method based on adaptive deentanglement in this invention. Detailed Implementation

[0062] The present invention will be further described in detail below with reference to embodiments:

[0063] like Figure 1 and Figure 2 As shown, this invention provides a cross-domain recommendation method based on adaptive disentanglement, comprising the following steps:

[0064] S1. Data Acquisition: Acquire historical interaction data from the source and target domains, analyze and select overlapping users in the source and target domains, and extract corresponding user-item interaction data from the source and target domains based on these overlapping users; where overlapping users refer to users who have appeared in both the source and target domains.

[0065] S2. Constructing the User-Item Interaction Graph: Using the overlapping user and item interaction data from the source and target domains, construct user-item interaction graphs for the source, target, and shared domains. The user-item interaction graph is a heterogeneous bipartite graph, meaning it includes user type nodes and item type nodes. The interaction graphs for the source and target domains use users and items from their respective domains, while the interaction graph for the shared domain includes users and items from both the source and target domains. In other words, for the interaction graphs of the source and target domains, users are overlapping users, and items are items from their respective domains. For the interaction graph of the shared domain, users are overlapping users, and items consist of items from both the source and target domains.

[0066] S3. User and item representation initialization: User and item representations are initialized separately for the source and target domains, respectively. For the shared domain, only the user representation is initialized, while its item representation is combined with the item representations of the source and target domains.

[0067] S4. Variational Bidirectional Graph Coding (VBGE): VBGE is used to encode the interaction graphs of the source domain, target domain, and shared domain respectively. Information is propagated through VBGE to generate approximate posterior distributions of users and items. Latent variable representations of users and items are generated by summarizing homogeneous information, capturing knowledge between the source domain and the target domain.

[0068] VBGE encodes the user-project interaction graph through a two-step information propagation mechanism, specifically including the following steps:

[0069] S401, Homogeneous Information Propagation: Since the user-item interaction graph is a heterogeneous bipartite graph, that is, the user-item interaction graph includes user type nodes and item type nodes, VBGE adopts a two-step information propagation mechanism to propagate information between homogeneous neighbors in order to generate intermediate representations;

[0070] S402, Intermediate Representation Calculation: For the user representation of the source domain, calculate its intermediate representation based on 2-hop neighbors.

[0071]

[0072] Where Norm(·) is the row normalization operation, A X Let δ(·) be the source domain user-project interaction matrix, and let δ(·) be the Leaky ReLU activation function. The learnable parameter matrix;

[0073] S403, Latent Variable Distribution Modeling: Based on Intermediate Representation Calculate the distribution of latent variables for users:

[0074]

[0075] in, For the concatenation operation, φ(·) is the Softplus function. and For learnable parameter matrix, and These are the mean and standard deviation represented by the user, respectively;

[0076] S404, Latent Variable Sampling: Sampling the user's latent variable representation using the reparameterization trick.

[0077]

[0078] in, A source domain-specific representation for the user, used to capture the user's characteristics within that domain;

[0079] S5. Adaptive deentanglement: Adaptive deentanglement technique is used to separate the shared information and domain-specific information in the source domain user representation and the same user representation in the shared domain to obtain the shared information dimension of the source domain user.

[0080] Specifically, the entanglement of shareable information dimensions is de-entangled using information provided by users rather than individual users. Therefore, it is not necessary to specify the number or position of dimensions to adaptively de-entangle specific dimensional information, thereby achieving the effect of quantifying user-shared information. Calculations are performed using formulas (5) and (6), and the dimension containing the shared information is determined by a threshold.

[0081]

[0082] in, For the user's representation distribution, D KL To obtain the KL divergence operation, δ i The KL divergence is the distance between users across different dimensions, where τ is the threshold; through δ u <τ, which determines the dimension and number of dimensions of the user's shared information;

[0083] S6. Shared Information Enhancement: The shared information enhancement technique is used to enhance the user representation of the target domain by de-entanglement of the source domain and the shared domain, and at the same time, the user representation of the source domain is enhanced by de-entanglement of the target domain and the shared domain, thereby improving the recommendation effect of the target domain and the source domain.

[0084] Specifically, firstly, the mean of the representation distribution of users in the source or target domain and the corresponding representation distribution in the shared domain is calculated based on the determined shareable dimension to obtain the average user representation distribution. Then, the user representation distribution is resampled to obtain the specific representation of the shareable user information. Finally, the user information is combined with the corresponding user representation to enhance the information contained in the user representation. The process of enhancing shared information is described as shown in formulas (7) and (8):

[0085]

[0086] user_share=mask⊙z i (8)

[0087] Where μ is the mean of the distribution, σ is the standard deviation of the distribution, and ∈ i To obtain from the standard normal distribution The noise term in the mid-sample, z i This represents the result of sampling the distribution; mask indicates the shareable dimension, and user_share is the shareable information obtained after de-entanglement;

[0088] S7. Personalized Recommendation: The final user representation and item representation are multiplied by a dot product to determine the user's interest in the items. Personalized recommendations are then provided based on this interest. First, a weighted summation method is used to fuse the shared representation and domain-specific representation obtained from the de-entanglement of the user's data. After the user and item representations are generated, a dot product is used to measure the user's interest in the items. Then, the interest is normalized using a sigmoid activation function for non-linear transformation to ensure that the calculated interest is suitable for ranking recommendations. Finally, based on the normalized interest, all candidate items are ranked, and the top N items with the highest interest are selected as the recommendation results. Specifically, the following steps are included:

[0089] S701. Generation of User Representations: In cross-domain recommendation scenarios, user representations consist of shared representations and domain-specific representations, which are fused using a weighted summation method to fully utilize cross-domain information while maintaining domain personalization. The calculation formula for user representations is as follows:

[0090]

[0091] Among them, E u This represents the final user representation vector; This represents the decoupled user-shared representation, which represents users' common interest information across all domains; This represents the user's original personalized representation in the current target domain, indicating the user's specific preferences in that domain;

[0092] S702. Calculate user interest in the project: In user representation E u After generation, the corresponding project representation is obtained, and the dot product is used to measure the user's interest in the project:

[0093] E i =E[i] (10)

[0094] s(u,i)=E u ·E i (11)

[0095] in, The item representation matrix obtained from model training, where N is the total number of items, d is the embedding dimension, i is the item index, i.e., the ID of the target item; E i Let be the representation vector of item i; s(u,i) represents the degree of interest of user u in item i;

[0096] S703. Normalization: The interest scores are normalized to ensure that the calculated interest scores s(u,i) are suitable for ranking recommendation. A sigmoid activation function is used for non-linear transformation.

[0097]

[0098] This normalization process ensures that the value of interest is within the range of (0,1), making the score more stable.

[0099] S704, Personalized Recommendation: Based on normalized interest levels All candidate items are sorted, and the top N items with the highest interest are selected as the recommendation results, i.e.:

[0100]

[0101] Then, the top N ranked projects are recommended to the user to achieve personalized recommendations.

[0102] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A cross-domain recommendation method based on adaptive deentanglement, characterized in that... Includes the following steps: S1. Data Acquisition: Acquire historical interaction data of the source domain and target domain, analyze and select overlapping users of the source domain and target domain, and extract the corresponding user-item interaction data from the source domain and target domain respectively based on the above overlapping users. S2. Constructing a User-Item Interaction Graph: Using the overlapping user and item interaction data of the source domain and the target domain, construct user-item interaction graphs for the source domain, the target domain, and the shared domain. The user-item interaction graph is a heterogeneous bipartite graph, which includes user type nodes and item type nodes. The interaction graphs of the source domain and the target domain use users and items within their respective domains, and the interaction graph of the shared domain includes users and items from both the source domain and the target domain. S3. User and item representation initialization: User and item representations are initialized separately for the source and target domains, respectively. For the shared domain, only the user representation is initialized, and its item representation is combined with the item representations of the source and target domains. S4. Variational Bidirectional Graph Coding (VBGE): VBGE is used to encode the interaction graphs of the source domain, target domain, and shared domain respectively. Information is propagated through VBGE to generate approximate posterior distributions of users and items. Latent variable representations of users and items are generated by summarizing homogeneous information, capturing knowledge between the source and target domains. S5. Adaptive deentanglement: Adaptive deentanglement technique is used to separate the shared information and domain-specific information in the source domain user representation and the same user representation in the shared domain to obtain the shared information dimension of the source domain user. S6. Shared Information Enhancement: The shared information enhancement technique is used to enhance the user representation of the target domain by de-entanglement of the source domain and the shared domain, and at the same time, the user representation of the source domain is enhanced by de-entanglement of the target domain and the shared domain, thereby improving the recommendation effect of the target domain and the source domain. S7. Personalized Recommendation: The final user representation and item representation are multiplied by a dot product to determine the user's interest in the item, and personalized recommendations are provided to the user based on this interest. In step S2, for the interaction graphs of the source and target domains, the users are overlapping users, and the items are items within their respective domains; for the interaction graphs of the shared domain, the users are overlapping users, and the items consist of items from both the source and target domains. In step S4, VBGE encodes the user-project interaction graph through a two-step information propagation mechanism, including the following steps: S401, Homogeneous Information Propagation: Since the user-project interaction graph is a heterogeneous bipartite graph, VBGE adopts a two-step information propagation mechanism to propagate information between homogeneous neighbors in order to generate intermediate representations; S402, Intermediate Representation Calculation: For the user representation of the source domain, calculate its intermediate representation based on 2-hop neighbors. : （1） in, For the normalization operation, For the source domain user-project interaction matrix, The LeakyReLU activation function is used. The learnable parameter matrix; S403, Latent Variable Distribution Modeling: Based on Intermediate Representation Calculate the distribution of latent variables for users: （2） （3） in, For splicing operations, For the Softplus function, , , and For learnable parameter matrix, and These are the mean and standard deviation represented by the user, respectively; S404, Latent Variable Sampling: Using the reparameterization trick to sample the user's latent variable representation. （4） in, A source domain-specific representation for the user, used to capture the user's characteristics in that domain.

2. The cross-domain recommendation method based on adaptive disentanglement according to claim 1, characterized in that: In step S5, calculations are performed using formulas (5) and (6), and the dimension containing the shared information is determined by using a threshold: （5） （6） in, For the user's representation distribution, To obtain the KL divergence operation, For the KL divergence among users in various dimensions, For threshold; through Determine the dimension and number of dimensions where the user-shared information is located.

3. The cross-domain recommendation method based on adaptive disentanglement according to claim 2, characterized in that: In step S6, the average user representation distribution is first calculated based on the determined shareable dimension for the representation distribution of users in the source domain or target domain and the corresponding representation distribution in the shared domain. Then, the user representation distribution is resampled to obtain the specific representation of the shareable user information. Finally, the user information is combined with the corresponding user representation to enhance the information contained in the user representation.

4. The cross-domain recommendation method based on adaptive disentanglement according to claim 3, characterized in that: In step S6, the process of enhancing shared information is described as shown in formulas (7) and (8): （7） （8） in, The mean of the distribution. Let be the standard deviation of the distribution. To obtain from the standard normal distribution The noise term in the sampled data. This is the result of sampling the distribution; The mask is used to specify the shareable dimensions. To untangle the information that can be shared.

5. The cross-domain recommendation method based on adaptive disentanglement according to claim 4, characterized in that: In step S7, the shared representation and domain-specific representation de-entangled by the user are first fused using a weighted summation method. After the user representation and item representation are generated, the user's interest in the item is measured by dot product calculation. Then, the interest is normalized and a nonlinear transformation is performed using the sigmoid activation function to ensure that the calculated interest is suitable for ranking recommendation. Finally, based on the normalized interest, all candidate items are ranked, and the top N items with the highest interest are selected as the recommendation results.

6. The cross-domain recommendation method based on adaptive disentanglement according to claim 5, characterized in that: Step S7 includes the following steps: S701. Generation of User Representations: In cross-domain recommendation scenarios, user representations consist of shared representations and domain-specific representations, which are fused using a weighted summation method to fully utilize cross-domain information while maintaining domain personalization. The calculation formula for user representations is as follows: （9） in, This represents the final user representation vector; This represents the decoupled user-shared representation, which represents users' common interest information across all domains; This represents the user's original personalized representation in the current target domain, reflecting the user's specific preferences in that domain; S702. Calculate user interest in the project: based on user feedback... After generation, the corresponding project representation is obtained, and the dot product is used to measure the user's interest in the project: （10） （11） in, The item representation matrix obtained from model training. For the total number of projects, For the embedded dimension; For the project index, For the project The representation vector; Indicates user For the project interest level; S703. Normalization Processing: Normalize the interest scores to ensure that the calculated interest scores are accurate. Suitable for ranking and recommendation, it uses the sigmoid activation function for non-linear transformation: （12） The interest level should be within the range of (0,1); S704, Personalized Recommendation: Based on normalized interest levels All candidate items are sorted, and the top N items with the highest interest are selected as the recommendation results, i.e.: （13） Then, the top N ranked projects are recommended to the user to achieve personalized recommendations.

7. A cross-domain recommendation method based on adaptive disentanglement according to any one of claims 1-6, characterized in that: In step S1, overlapping users refer to users who have appeared in both the source domain and the target domain.

Images