A bidirectional cross-domain recommendation method based on linear gate interpolation and migration gain

Abstract

The application provides a bidirectional cross-domain recommendation method based on linear gate interpolation and migration gain. The method uses linear gate interpolation to complete the shared user feature embedding enhancement in the dual domain; uses a hybrid model to model the cross-domain migration information and generate collaborative preference representation; uses an independent model to model the single-domain interaction information and generate domain preference features; based on the relationship between the hybrid loss and the independent loss, the migration information gain is used to obtain a dynamic adaptive bidirectional migration strategy to generate the final user preference representation. By introducing the interest enhancement strategy, calculating the migration gain, obtaining the feasibility of cross-domain information migration, and supplementing the cross-domain collaborative information for the field with positive migration gain, without introducing data sparse field knowledge into the data rich field, the negative migration problem caused by the migration of sparse cross-domain features to the rich domain can be avoided, the quality of the dual-domain user preference representation and the model generalization ability are effectively improved, and the personalized service purpose in the cross-border cooperation scene is realized.

Description

Technical Field

[0001] This invention belongs to the field of bidirectional cross-domain recommendation technology, and particularly relates to a bidirectional cross-domain recommendation method based on linear gated interpolation and migration gain. Background Technology

[0002] With the rapid development of computer technology and the Internet, more and more user interactions are no longer limited to the same domain, and cross-domain recommendation has become one of the important application technologies in the field of personalized recommendation services. However, due to the development requirements of refined service strategies for multi-domain interoperability and cooperation, traditional cross-domain recommendation technology can no longer meet the requirements of personalized recommendation services in cross-industry cooperation scenarios. Two-way cross-domain recommendation technology has emerged, which can link information from two domains to provide users with more accurate recommendations.

[0003] Cross-domain preference decoupling technology has the advantage of noise reduction. Most existing research results on bidirectional cross-domain recommendation achieve bidirectional cross-domain user preference construction by decoupling domain-shared features and domain-independent features, and have shown good recommendation performance on multiple datasets.

[0004] In 2023, researchers at Macquarie University proposed a fine-grained user preference decoupling method. It first uses a multilayer perceptron to learn a general representation of cross-domain preferences, and then employs reparameter sampling to extract domain-independent and domain-unique representations respectively. While fine-grained preference extraction techniques have shown good performance in bidirectional cross-domain recommendation, these methods suffer from the following problems: 1. Existing bidirectional cross-domain recommendation methods supplement current domain information with cross-domain information, neglecting the negative transfer problem that arises when sparse cross-domain features are transferred to richer domains; 2. In bidirectional cross-domain recommendation scenarios, there are relatively few shared users, resulting in a limited number of shared features, leading to poor fit and insufficient user preference fitting in the preference representation. In summary, existing bidirectional cross-domain methods learn poor-quality user preference representations and provide recommendation results that do not meet users' personalized needs, thus reducing the performance of the recommendation system. Summary of the Invention

[0005] This invention aims to provide a bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain, in order to overcome the problems of poor user preference representation quality and unsatisfactory recommendation results in existing technologies.

[0006] To achieve the objectives of this invention, the technical solution provided by this invention is: a bidirectional cross-domain recommendation method based on linear gated interpolation and migration gain, comprising the following steps:

[0007] Step 1: Obtain user interaction records for each domain, and embed features for users and items in each domain:

[0008] Step 2: Expand cross-domain sharing user interests using gating interpolation technology:

[0009] Step 3: Based on domain interaction information, construct user interest representations using a linear attention network:

[0010] Step 4: Calculate the losses of the mixture model and the independent model, and establish a bidirectional cross-domain transfer paradigm based on the transfer gain relationship:

[0011] 4.1 Based on cross-domain user interests, enhanced user interests, and single-domain user interaction records, a hybrid model is used to establish a cross-domain migration collaborative preference representation:

[0012] 4.2. Based on user embedded representations and user interaction records, construct domain preference representations using independent models:

[0013] 4.3 Calculation of transfer information gain: Calculate the transfer information gain formed by the collaborative preference representation and the domain user preference representation in each domain to obtain the domain feature weights for bidirectional cross-domain learning.

[0014] Step 5: Calculate the interaction probability for all candidate nodes and generate recommendation lists for different domains simultaneously.

[0015] Furthermore, the calculation in step two above is as follows:

[0016]

[0017]

[0018] in For user u j Embedded representation in domain A For user u j Embedded representation in the B domain For user u j Enhanced embedding representation in domain A For user u j The enhanced embedding representation in the B domain, λ∈[0,1] and follows a Beta distribution.

[0019] Furthermore, in section 4.1 above, the loss function is:

[0020]

[0021]

[0022] Where f(·) is the interest fusion model, g(·) is the bidirectional cross-domain transfer model, and θ mix(A) ,θ mix(B) These are the parameters for model training.

[0023] Furthermore, in section 4.2 above, the loss function is:

[0024]

[0025]

[0026] Furthermore, in section 4.3 above, the calculation of migration information gain includes the following steps:

[0027] (1) The weighted average of the mixed loss of transfer learning and the independent loss of domain encoding is calculated using the following formula:

[0028]

[0029]

[0030] (2) When the gain value is greater than 0, the user preference is represented as a weighted sum of the independent model and the mixed model, which is expressed as:

[0031] When Δ A >0,Δ B When >0,

[0032]

[0033] When Δ A =0,Δ B When = 0,

[0034]

[0035] When Δ A =0,Δ B When >0,

[0036]

[0037] When Δ A >0,Δ B When = 0,

[0038]

[0039] Compared with the prior art, the advantages of the present invention are:

[0040] 1. The method of this invention first establishes a cross-domain shared user interest enhancement mechanism based on a linear gated interpolation function by learning the interest features of shared users in various domains. Secondly, based on the user interest change characteristics in a single domain, it captures bidirectional transfer information gain by calculating the loss of a dual-domain hybrid model and the loss of a single-domain independent model. Then, according to the amount of transfer information gain, it dynamically and adaptively establishes the weights for bidirectional transfer learning, enabling the final bidirectional cross-domain transfer model to have autonomous selectivity, thereby generating a high-quality user preference representation. Through the mining of association diversity features and the modeling of the dynamically adaptive bidirectional cross-domain transfer paradigm, the final user preference representation has strong representational ability and a sufficient fit to user preferences.

[0041] 2. This invention targets bidirectional cross-domain recommendation scenarios. By designing interest enhancement strategies and adaptive cross-domain mechanisms, it effectively models interest transfer patterns and constructs an adaptive bidirectional cross-domain transfer paradigm. This alleviates the problems of insufficient user preference fitting and negative transfer in cross-domain recommendation, greatly improves the quality of user preference fitting in both domains and the generalization ability of the bidirectional cross-domain recommendation model, and has the advantage of improving the quality of user preference representation. In turn, it enhances the effect of personalized recommendation, thereby achieving the goal of personalized service in cross-domain cooperation scenarios.

[0042] Figure label:

[0043] Figure 1 This is a structural diagram of the method of the present invention;

[0044] Figure 2 This is a flowchart of the present invention. Detailed implementation method:

[0045] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0046] The design concept of this invention is as follows: First, a linear gated interpolation method is used to combine single-domain interests and cross-domain interests to form an enhanced user interest embedding. Second, a linear attention network is used to model single-domain user interaction relationships to form a domain user preference representation. Then, a hybrid model is used to obtain a cross-domain transfer collaborative preference representation, while an independent model is used to obtain domain preference features based on single-domain user preferences, enhanced user interests, and interaction relationships. On this basis, transfer information gain technology is used to calculate the weighting between the hybrid loss generated by generative transfer learning and the independent loss generated by domain encoding, and an adaptive bidirectional cross-domain transfer paradigm is constructed. Finally, combining the candidate item representations in each domain, similarity correlation is used to calculate the probability of items that the user may interact with in the next time step, and the next interactive item is generated accordingly, thus obtaining a bidirectional cross-domain recommendation result.

[0047] Based on the above design concept, this invention provides a bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain, such as... Figure 1 As shown, the specific steps include:

[0048] Step 1: Obtain user interaction records for each domain, and embed features for users and items in each domain:

[0049] For bidirectional cross-domain recommendation scenarios, V is adopted respectively. A V B Represents a unique set of projects in domains A and B, using U A U B Let A represent the unique set of users in domains A and B. Given single-domain user interaction information, construct multilayer perceptron encoders in domains A and B respectively to map one-hot encoded features in each domain to low-dimensional dense spatial features. The users in domain A... and projects Feature embeddings are respectively represented as and n represents the feature embedding dimension in domain A; and the user in domain B. and projects Feature embeddings are respectively represented as and m represents the feature embedding dimension in the B domain.

[0050] Step 2: Expand cross-domain sharing user interests using gating interpolation technology:

[0051] For shared users in bidirectional cross-domain scenarios, linear gated interpolation is used to enhance the features of the shared users' interests. The calculation process is as follows:

[0052]

[0053]

[0054] in For user u j Embedded representation in domain A For user u j Embedded representation in the B domain For user u j Enhanced embedding representation in domain A For user u j The enhanced embedding representation in the B domain, λ∈[0,1] and follows a Beta distribution.

[0055] Step 3: Based on domain interaction information, construct user interest representations using a linear attention network:

[0056] For any user's interaction history, the item features are embedded into a linear attention network. The linear attention network independently encodes interactions in domain A and domain B, respectively, to obtain the user's interest representation h in each domain. A and h B The calculation process is expressed as follows:

[0057]

[0058] LayerNorm(·) is the normalization function, and MH(·) is the multi-head attention module.

[0059] Step 4: Calculate the losses of the mixture model and the independent model, and establish a bidirectional cross-domain transfer paradigm based on the transfer gain relationship:

[0060] 4.1 Based on cross-domain user interests, enhanced user interests, and single-domain user interaction records, a hybrid model is used to establish a cross-domain migration collaborative preference representation:

[0061] By transferring user interest features from domain B to domain A using a hybrid model, we obtain the collaborative preference representation h of domain A. AB Simultaneously, by using a hybrid model, interest features from domain A are transferred to domain B to obtain a collaborative preference representation h in domain B. BA Cross-domain user interest modeling can employ any bidirectional cross-domain transfer model. The loss function obtained by encoding the cross-domain transfer collaborative preference representation through a hybrid model is:

[0062]

[0063]

[0064] Where f(·) is the interest fusion model, g(·) is the bidirectional cross-domain transfer model, and θ mix(A) ,θ mix(B) These are the parameters for model training.

[0065] 4.2. Based on user embedded representations and user interaction records, construct domain preference representations using independent models:

[0066] For user interest characteristics in various fields, h A and h B As input to the independent model, the loss function is obtained by encoding the user domain preference representation:

[0067]

[0068]

[0069] 4.3 Calculation of Transfer Information Gain:

[0070] In this step, drawing on the concept of information gain, the transfer information gain formed by the collaborative preference representation and the domain user preference representation in each domain is calculated to obtain the domain feature weights for bidirectional cross-domain learning.

[0071] (1) To establish the influence relationship of bidirectional cross-domain information, the transfer information gain is used to generate the weighted average of the mixed loss of transfer learning and the independent loss of domain encoding. The calculation formula is as follows:

[0072]

[0073]

[0074] (2) Based on the migration information gain Δ A and Δ B When the gain value is less than or equal to 0, it indicates that introducing cross-domain information will lead to negative transfer in bidirectional cross-domain recommendation, resulting in a decrease in the quality of user preference learning. Conversely, it indicates that introducing cross-domain information is effective. Therefore, when the gain value is greater than 0, the final user preference is represented as a weighted sum of the independent model and the hybrid model. This process is expressed as:

[0075] When Δ A >0,Δ B When >0,

[0076]

[0077] When Δ A =0,Δ B When = 0,

[0078]

[0079] When Δ A =0,Δ B When >0,

[0080]

[0081] When Δ A >0,Δ B When = 0,

[0082]

[0083] Step 5: Calculate the interaction probability for all candidate nodes and simultaneously generate recommendation lists for different domains:

[0084] In each domain, the user's next interaction item is determined by calculating the probability distribution of the user's preference representation and all candidate items in each domain. The specific calculation process is as follows:

[0085]

[0086] in It is the user in domain A interacting with project v j The probability of interacting in the next moment. It is the user's interaction with the project in domain B. j The probability of interaction at the next moment is calculated. Then, based on the probability calculation results, the items are sorted from largest to smallest, and the items at the top of the list are more likely to interact with the user.

[0087] The model uses a cross-loss function to calculate the difference between the predicted and true values. Therefore, for user interaction relationships in domain A, the loss function formula is:

[0088]

[0089] Similarly, for user interaction relationships in domain B, the loss function calculation formula is:

[0090]

[0091] Combining the mixed loss function and the independent loss function, the final loss function of the model is:

[0092] Loss = L A +L B +Loss mix(A) +Loss mix(B) +Loss A +Loss B

[0093] This invention introduces an interest enhancement strategy to calculate the migration gain, obtain the feasibility of cross-domain information migration, and supplement the domain with positive migration gain through cross-domain collaborative information, without introducing sparse domain knowledge into the data-rich domain. This avoids the negative migration problem that occurs when sparse cross-domain features migrate to the rich domain, effectively improving the quality of dual-domain user preference representation and model generalization ability, thereby achieving the goal of personalized service in cross-domain cooperation scenarios.

[0094] The above description is a specific illustration of the present invention, and not a limitation thereof. Those skilled in the art can make various equivalent technical solutions without departing from the scope of the present invention; therefore, all equivalent technical solutions should fall within the patent protection scope of the present invention.

Claims

1. A bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain, characterized in that, Includes the following steps: Step 1: Obtain user interaction records for each domain, and embed features for users and items in each domain: Step 2: Expand cross-domain sharing user interests using gating interpolation technology: Step 3: Based on domain interaction information, construct user interest representations using a linear attention network: Step 4, Migration Gain: Step 4.1 Based on cross-domain user interests, enhanced user interests, and single-domain user interaction records, a hybrid model is used to establish a cross-domain migration collaborative preference representation: Step 4.2: Based on user embedded representations and user interaction records, construct domain preference representations using independent models: Step 4.3, Calculation of transfer information gain: Calculate the transfer information gain formed by the collaborative preference representation and the domain user preference representation in each domain to obtain the domain feature weights for bidirectional cross-domain learning; Step 5: Calculate the interaction probability for all candidate nodes and generate recommendation lists for different domains simultaneously.

2. The bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain according to claim 1, characterized in that, The calculation in step two is as follows: in For users Embedded representation in domain A For users Embedded representation in the B domain For users Enhanced embedding representation in domain A For users Enhanced embedding representation in the B domain And it follows a Beta distribution.

3. The bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain according to claim 2, characterized in that, In step 4.1, the loss function is: in For interest fusion model, This is a two-way cross-domain migration model. For model training parameters, This represents the user's interests in domain A. This represents the user's interests in domain B. For the representation of collaborative preferences in domain A, This represents the collaborative preference representation in domain B.

4. The bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain according to claim 3, characterized in that, In step 4.2, the loss function is: 。 5. The bidirectional cross-domain recommendation method based on linear gated interpolation and transfer gain according to claim 4, characterized in that, Step 4.3, the calculation of migration information gain, includes the following steps: (1) The weighting of the mixed loss of transfer learning and the independent loss of domain encoding is calculated using the following formula: (2) When the gain value is greater than 0, the user preference is represented as a weighted sum of the independent model and the mixed model, which is expressed as: when hour, when hour, when hour, when hour, 。

Images