A universal cross-domain recommendation method based on preference prototype-aware learning

Abstract

The present invention discloses a general recommendation method based on preference prototype-aware learning, which belongs to big data analysis technology. When realizing general cross-domain recommendation, the present invention quantitatively learns user preferences through a preference prototype-aware learning method while minimizing interference from the original domain. The method of the present invention consists of two complementary components: a hybrid encoder and a preference prototype-aware decoder, forming an end-to-end unified framework suitable for various real-world scenarios. The hybrid encoder uses a hybrid network to learn a general representation of interacting items and capture the intrinsic relationship between items across different domains. The preference prototype-aware decoder implements a learnable prototype matching mechanism to quantitatively perceive user preferences, which can accurately capture user preferences at a higher semantic level. The preference prototype-aware decoder can also avoid interference caused by item features of the original domain. The advantage of the present invention lies not only in providing accurate recommendations, but also in providing reliable preference prototypes.

Description

Technical Field

[0001] The present invention belongs to the field of big data analysis technology, specifically to information recommendation technology, and in particular to a universal cross-domain recommendation method based on preference prototype perception learning. Background Art

[0002] Recommendation technology aims to provide users with satisfying information push services and is widely used in numerous online scenarios, such as friend recommendations on social media platforms, product recommendations on shopping platforms, and video recommendations on short video platforms. Recommendation systems have revolutionized how users discover and engage with content, primarily by uncovering potential user preferences and suggesting items of potential interest. Single-domain recommendation systems focus on modeling user behavior within a single domain, generating recommendations by analyzing historical user behavior and item characteristics. The systems continuously optimize the recommendation results, providing personalized recommendations tailored to user preferences and needs.

[0003] Cross-domain recommendation aims to enhance the understanding of user preferences by leveraging data from multiple domains, thereby achieving more personalized and accurate recommendations. Traditional recommender systems primarily focus on dual-domain research and lack dedicated modules for modeling user preferences. When attempting to extract user preferences across multiple domains, these methods face challenges in identifying true user preferences and filtering redundant item features, resulting in suboptimal performance in this context. To address these issues, some methods have constructed specialized modules to extract user preferences. These methods, based on a unified framework, are easily applicable in real-world dual-domain and multi-domain scenarios and are therefore categorized as general cross-domain recommendation methods. This pioneering work extracts item-specific features through an aggregator consisting of simple pooling or attention mechanisms. Inspired by this general and effective framework, some methods have utilized fine-tuning techniques to re-transfer user preferences. However, these methods require pre-training, fall short of an end-to-end architecture, and are unable to directly extract user preferences. Recently, some methods have masked the extracted user preferences by item representations, aiming to encourage the model to indirectly predict user preferences. A key limitation of this approach is that the masking mechanism is particularly challenging for datasets with limited data.

[0004] Despite promising results, the aforementioned cross-domain recommendation research primarily focuses on extracting user preferences and specific features from the source domain. However, this approach may not effectively capture the true nature of user preferences, as user preferences are more closely associated with items interacting at a higher semantic level rather than with specific item features. Overemphasizing item-specific features can lead to suboptimal recommendations and hinder the transfer of user preferences across domains. Considering only specific book features, such as author, genre, or writing style, can lead to an overly narrow focus on book-related features. This item-feature-centric approach can make it difficult to recommend relevant items from other domains (such as movies or TV shows) that align with user preferences, especially when these items are related to the user's comedy preferences. This reliance on item-specific features can introduce noise and limit the generalization ability of recommendation systems. All of these methods utilize item-specific features to roughly represent user preferences, while ignoring the significant noise introduced by source domain features and irrelevant specific features. This neglect ultimately leads to suboptimal recommendations based on redundant specific features, as the model cannot effectively distinguish between valid and irrelevant features. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology in extracting redundant features, the present invention proposes a universal recommendation method based on preference-aware learning, which can better learn the general representation of interactive items, capture the intrinsic relationship between items in different fields, and realize the accurate extraction of user preferences, thereby achieving better recommendation effects.

[0006] A universal cross-domain recommendation method based on preference prototype-aware learning is proposed. The method utilizes prototype-aware learning to quantify user preferences and implements universal cross-domain recommendations in multiple scenarios based on given specific interaction items, global interaction items, and user IDs. The method includes the following steps:

[0007] S1. Define the users, items, and interactions in the domain. The dual-domain scenario composed of domain X and domain Y is represented as i∈{X,Y}, and the set of users, items, and interactions is represented as Similarity score B between cross-domain projects g =(B X ,B Y ), where B X and B Y Represent the personalized similarity scores in domain X and domain Y, respectively. The similarity scores are used to enhance the prior knowledge of item features;

[0008] Based on interactive ε i From the project Select N items to form a specific interaction item Based on interactive ε i from and Select 2N items to represent shared interactive items V g ,and and B i and global embedding V g and B g The relevant inputs that constitute the project effectiveness;

[0009] S2. We construct a hybrid encoder by extracting effective specific representations guided by item ratings, and learn a universal representation of interacted items based on the hybrid network. We capture the domain characteristics of the two encodings based on two hybrid encoder branches, capturing deeper dependencies between items, and then generate learnable weights through item ratings to guide item embedding.

[0010] S3. Based on the relationship between the items captured by the hybrid network in step S2, the specific branch and the global branch are encoded to make them highly correlated specific domain features and global domain features, and then the hybrid encoder is used to learn the representation of the specific domain and the global domain, and the encoded features of the specific branch and the global branch are included for use by the prototype perception decoder;

[0011] S4. Use the user interest prototype quantified by the prototype decoder to locate the target project, including quantifying the user preferences in the source domain from dislike to like according to positive and negative preferences, thereby obtaining the prototype preference representation, calculating the project similarity based on the prototype-aware decoder, introducing the preferences of the source domain into the shared domain to guide the global encoding features, and extracting the final user features in the target domain. Finally, the user-item prediction score is passed Calculation to determine recommended items;

[0012] The method involves selecting contrastive learning as an auxiliary task algorithm to condition the latent space of positive and negative interactions.

[0013] Based on the above solution, step S1 further includes the following steps:

[0014] definition and Represents the embedding of users and items and is used to extract effective representations, where h is the global embedding dimension and u represents the data user embedding. Based on this embedding layer, three basic representations that can be applied to cross-domain recommendations in multi-target scenarios are obtained: and (V g ,B g ), u represents user embedding, Indicates a score of B i , with a score of B g Global project embedding v g .

[0015] Furthermore, in step S2, potential user u is considered j Interactions in domain X are embedded in the context of and score B X ∈R N×1 Based on the historical records, a hybrid network is designed to capture the domain representation E for compressing item embedding and scoring, where N is the number of samples;

[0016] The hybrid network encodes item embeddings via parameter ψ, denoted as f(·; ψ): R N×h →R 1×h , the mixing mechanism followed by this mixing network is as follows:

[0017]

[0018] in, and is a learnable parameter, the symbol |·| represents the absolute value, and the mixing weight W1∈R N×h and W2∈R h×1 is learned from the input score B; after running the above formula, through the domain representation E and the current user embedding u j , generate the final encoded domain features

[0019] Furthermore, in step S3, the specific branch is to is provided to the hybrid encoder to capture the underlying specific features, which is represented as follows:

[0020]

[0021] The global branch refers to the input [u j ,(v g ,B g )] is provided to the hybrid encoder to generate the global encoding features, which are expressed as follows:

[0022]

[0023] In this step, the encoding features Will be further used by the decoder.

[0024] Furthermore, the step S4 quantifies the preferences of users in the source domain by dividing the positive preferences into K adjustable prototypes and dividing the negative preferences into K prototypes to quantify the preferences from dislike to like;

[0025] The above prototype is represented as User preferences are extracted from the source domain, so they are only relevant to interaction, so by matching the prototype p with the highest similarity * To make the model aware of preferences, the similarity is calculated as follows:

[0026]

[0027] Among them, ϵ is set to a small value to prevent division by zero. * Finally, the global encoding features are guided by introducing this preference of the source domain into the shared domain Thus effectively extracting the final user features in the target domain The linear layer g u With parameter φ, it is used to dynamically adapt the encoded features to the changing preference features and extract the final user features:

[0028]

[0029] Objective function of prototype learning The definition is as follows:

[0030]

[0031] Among them, c is a fully connected layer used to predict the probability of positive or negative, g p is the prototype layer, represents the cross entropy loss for preference classification within the prototype branch; l is an array used to distinguish item categories, and λ1, λ2, and λ3 are weights for controlling the loss;

[0032] To build the final prototype, the constraints considered included:

[0033] (1) Setting clustering loss To encourage these interaction items in the source domain to approach a prototype corresponding to their preferences;

[0034]

[0035] (2) Setting separation loss to promote an increase in the distance between the characteristics of the coded item and the prototype that does not belong to its preference;

[0036]

[0037] (3) Setting diversity loss By penalizing prototypes that are too close, the diversity of the learned prototypes is encouraged;

[0038]

[0039] In the above formula, It is y iThe prototype set under the class, k = {1, 2} represents two preferences, ξ is the threshold of cosine similarity in diversity loss;

[0040] Finally, the user's corresponding project prediction score is passed Calculation to determine recommended projects;

[0041] This method also uses binary cross entropy loss to predict the loss:

[0042]

[0043] in, Indicates uninteresting items by dot product calculation Negative prediction score.

[0044] Furthermore, the specific steps of using contrastive learning as an auxiliary task algorithm to adjust the potential space of positive and negative interactions are as follows:

[0045] Assume that on domain K, we are entering a specific project on the master branch Same, negative inputs are items of no interest Random sampling in; after the main branch, get positive features and negative features The auxiliary losses are as follows:

[0046]

[0047] Where, represents the cross entropy loss of classification, Z p is a positive label, indicating 1, Z n is a negative label, indicating 0;

[0048] The model constructed by the method is To optimize:

[0049]

[0050] Beneficial effects: The general cross-domain recommendation method based on preference prototype perception learning provided by the present invention has the following significant effects:

[0051] (1) This paper redesigns the extraction of user preferences and studies the redundant item feature interference problem caused by the item-centric extraction in the source domain, which can capture the real user preferences more accurately.

[0052] (2) The present invention mines user preferences from quantitative prototypes, and can minimize the impact of project interference while learning accurate user preferences.

[0053] (3) By integrating a hybrid encoder and a prototype decoder, the present invention can effectively capture user preferences in various scenarios in a unified end-to-end framework. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1(a) is a schematic diagram of the process structure of the traditional method;

[0055] Figure 1(b) is a schematic diagram of the framework structure of the method of the present invention;

[0056] Figure 2 It is a schematic diagram of the overall structure of the model of the present invention. DETAILED DESCRIPTION

[0057] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0058] This invention provides a general recommendation method based on preference-aware learning, which aims to better learn the general representation of interactive items, capture the intrinsic relationships between items in different fields, and accurately extract user preferences, thereby achieving better recommendation results. The implementation steps of this method are as follows:

[0059] Step 1: Define users, projects, and interactions in the domain.

[0060] For a dual-domain scenario i∈{X,Y}, the recommended data consists of users, items, and interactions. For the interaction ε i The present invention is achieved by B g =(B X ,B Y ) Calculate the preprocessed item-item similarity. represents the individualized item score in domain X, and B Y represents the personalized item scores in domain Y. These similarity scores are used to enhance the prior knowledge of item characteristics.

[0061] The present invention is based on i from Select N items to represent specific interactive items The present invention is based on i From v X and Select 2N items to represent shared interactive items V g . and B i and V g and B g In order to further extract its effective representation, the present invention introduces and to represent the embedding of users and items, where h is the global embedding dimension. After these embedding layers, three basic representations for cross-domain recommendation are obtained: user embedding u, with rating B i Specific project embedding and has a rating of B g Global project embedding v g The present invention introduces the following into the multi-target scenario: and (V g ,B g ), and B i and V g and B g constitutes valid project-related input.

[0062] Step 2: Extract effective specific representations and construct a hybrid encoder guided by item ratings. Unlike traditional simple multiplication, this paper generates learnable weights through ratings to guide item embedding.

[0063] In step 2, potential user u j Interactions occurred in domain X, and the corresponding history includes specific item embeddings and score B X ∈R N×1 , where N is the number of samples. To capture domain representation, we design a hybrid network to compress item embeddings and ratings.

[0064] The present invention uses a hybrid network f(·; ψ):R parameterized as ψ N×h →R 1×h , to encode item embeddings, following the general mixing mechanism as follows:

[0065]

[0066] in, and Is a learnable parameter. The symbol |·| indicates the absolute value. The mixing weight W1∈R N×h and W2∈R h×1 It is learned from the input score B.

[0067] After running the above formula, the domain representation E and the current user embedding u j , generate the final encoded domain features Here, the present invention adopts two hybrid encoder branches to capture the domain features of two codes.

[0068] Step 3: Generate different encoding features in the specific branch and the global branch.

[0069] In step 3, in a specific branch, enter Provided to a specific hybrid encoder, the focus is on capturing potential specific features. The specific encoding features generated by the hybrid encoder are as follows:

[0070]

[0071] In step 3, in the global branch, the input [u j ,(v g ,B g )] is provided to the hybrid encoder to generate the global encoding features as follows:

[0072]

[0073] In step 3, the present invention uses a hybrid encoder to learn representations of specific domains and global domains. The hybrid network can capture deeper dependencies between items and further encode them into highly correlated specific domain features and global domain features. These encoded features Will be further used by the decoder.

[0074] Step 4: Avoid relying on irrelevant item features in the source domain and further focus on potential user preferences in the global domain. This is achieved by using prototypes to represent preferences rather than directly using item features. At the core of the prototype-aware decoder are quantified user interest prototypes, which are used to locate target items and enable the decoder to clearly perceive user preferences.

[0075] In step 4, the present invention divides the positive preference into K adjustable prototypes and divides the negative preference into K prototypes, with a total of 2K prototypes, thereby accurately quantifying the preference from dislike to like. The present invention represents these prototypes as User preferences are extracted from the source domain, so they are only relevant to Interaction. By matching the prototype p with the highest similarity * To make the model aware of preferences, the specific similarity calculation is as follows:

[0076]

[0077] Among them, ϵ is set to a small value to prevent division by zero. * Finally, the global encoding features are guided by introducing this preference of the source domain into the shared domain Thus effectively extracting the final user features in the target domain Linear layer g u With parameter φ, it is used to dynamically adapt the encoded features to the changing preference features and extract the final user features:

[0078]

[0079] Objective function of prototype learning The definition is as follows:

[0080]

[0081] Among them, c is a fully connected layer used to predict the probability of positive or negative, g p It is the prototype layer. Represents the cross entropy loss for preference classification within the prototype branch. When inputting normal interaction items, l is an array of all 1s. λ1, λ2, and λ3 control the weights of the loss, and the present invention can follow the general values in classification tasks.

[0082] As shown in the above formula, in order to construct the final prototype, the present invention establishes the following constraints.

[0083] Clustering loss These interacting items in the source domain are encouraged to approach a prototype corresponding to their preference. Promotes an increase in the distance between the characteristics of the encoded item and the prototype that does not belong to its preference.

[0084]

[0085] in, yes The prototype set under the class. In addition, the diversity loss in the above formula By penalizing prototypes that are too close, diversity in the learned prototypes is encouraged.

[0086]

[0087] Where k = {1, 2} represents two preferences and ξ is the threshold of cosine similarity in diversity loss.

[0088] Finally, the user-item prediction score is obtained by Calculation to decide which item to recommend.

[0089] This paper follows previous work and uses binary cross entropy loss for basic prediction loss:

[0090]

[0091] in, Indicates uninteresting items by dot product calculation Negative prediction score.

[0092] Step 5: The present invention selects contrastive learning as an auxiliary task algorithm to adjust the potential space of positive and negative interactions.

[0093] Assume that on domain K, we are entering a specific project on the master branch Negative inputs are items of no interest. After the main branch, we get the positive features and negative features The auxiliary losses are as follows:

[0094]

[0095] in, represents the cross entropy loss for classification. Z p is a positive label, indicating 1. Z n is a negative label, indicating 0. The entire model corresponding to the present invention can be represented by the following overall objective function To optimize:

[0096]

[0097] Example

[0098] In order to verify the effectiveness of this method, the present invention conducts experiments on data sets corresponding to four different scenarios to make them meet the experimental conditions of the present invention.

[0099] For the product recommendation task, this paper adopts HitRatio (HR) and Normalized Discounted Cumulative Gain (NDCG) as evaluation criteria.

[0100] The present invention compares the effects with different comparison methods in four recommendation scenarios (within two domains, between two domains, within multiple domains with overlapping items, and within multiple domains with overlapping users). These four methods cover the vast majority of representative recommendation scenarios in general recommendation, proving the versatility of the model proposed in this invention.

[0101] Table 1. Comparison of recommendation performance in dual-domain recommendation scenarios

[0102]

[0103] Table 2. Comparison of recommendation performance in dual-domain recommendation scenarios

[0104]

[0105] Table 3. Comparison of recommendation performance in the scenario of overlapping items and multiple domains

[0106]

[0107] Table 4. Comparison of recommendation performance in multi-domain recommendation scenarios with overlapping users

[0108]

[0109] Specifically, Tables 1, 2, 3, and 4 respectively show the experimental results in four representative recommendation scenarios for general recommendation. It can be seen that in these four scenarios, the method proposed in the present invention performs well in both HR and NDCG indicators, demonstrating superior performance in the field of general recommendation.

Claims

1. A general cross-domain recommendation method based on preference prototype perception learning, characterized by: The method includes using prototype-aware learning to quantify user preferences and, based on given specific interaction items and global interaction items and user IDs, to achieve universal cross-domain recommendations in several scenarios. The method steps include: S1. Define the users, items, and interactions in the domain. The dual-domain scenario composed of domain X and domain Y is represented as i∈{X,Y}, and the set of users, items, and interactions is represented as Similarity score B between cross-domain projects g =(B X ,B Y ), where B X and B Y Represent the personalized similarity scores in domain X and domain Y, respectively. The similarity scores are used to enhance the prior knowledge of item features; Based on interaction From the project Select N items to form a specific interaction item Based on interaction from and Select 2N items to represent shared interactive items V g ,and and B i and global embedding V g and B g The relevant inputs that constitute the project effectiveness; S2. We construct a hybrid encoder by extracting effective specific representations guided by item ratings, and learn a universal representation of interacted items based on the hybrid network. We capture the domain characteristics of the two encodings based on two hybrid encoder branches, capturing deeper dependencies between items, and then generate learnable weights through item ratings to guide item embedding. Consider potential user u j Interactions in domain X are embedded in the context of and score B X ∈R N×1 Based on the historical records, a hybrid network is designed to capture the domain representation E for compressing item embedding and scoring, where N is the number of samples; The hybrid network encodes item embeddings via parameter ψ, denoted as f(·; ψ): R N×h →R 1×h , the mixing mechanism followed by this mixing network is as follows: in, and is a learnable parameter, the symbol |·| represents the absolute value, and the mixing weight W1∈R N×h and W2∈R h ×1 is learned from the input score B; After running the above formula, the domain representation E and the current user embedding u j , generate the final encoded domain features S3. Based on the relationship between the items captured by the hybrid network in step S2, the specific branch and the global branch are encoded to make them highly correlated specific domain features and global domain features, and then the hybrid encoder is used to learn the representation of the specific domain and the global domain, and the encoded features of the specific branch and the global branch are included for use by the prototype perception decoder; S4. Use the user interest prototype quantified by the prototype decoder to locate the target project, including quantifying the user preferences in the source domain from dislike to like according to positive and negative preferences, thereby obtaining the prototype preference representation, calculating the project similarity based on the prototype-aware decoder, introducing the preferences of the source domain into the shared domain to guide the global encoding features, and extracting the final user features in the target domain. Finally, the user-item prediction score is passed Calculation to determine recommended items; The method involves selecting contrastive learning as an auxiliary task algorithm to condition the latent space of positive and negative interactions.

2. The universal cross-domain recommendation method based on preference prototype-aware learning according to claim 1 is characterized by: Step S1 also includes the following steps: definition and Represents the embedding of users and items and is used to extract effective representations, where h is the global embedding dimension and u represents the data user embedding. Based on this embedding, three basic representations that can be applied to cross-domain recommendations in multi-target scenarios are obtained: u, and (V g ,B g ), u represents user embedding, Indicates a score of B i , with a score of B g Global project embedding v g .

3. The universal cross-domain recommendation method based on preference prototype perception learning according to claim 1 is characterized by: In step S3, the specific branch is to is provided to the hybrid encoder to capture the underlying specific features, which is represented as follows: The global branch refers to the input [u j ,(v g ,B g )] is provided to the hybrid encoder to generate the global encoding features, which are expressed as follows: In this step, the encoding features Will be further used by the decoder.

4. The universal cross-domain recommendation method based on preference prototype-aware learning according to claim 1 is characterized by: The step S4 quantifies the preferences of users in the source domain by dividing the positive preferences into K adjustable prototypes and dividing the negative preferences into K prototypes to quantify the preferences from dislike to like; The above prototype is represented as User preferences are extracted from the source domain, so they are only relevant to interaction, so by matching the prototype p with the highest similarity * To make the model aware of preferences, the similarity is calculated as follows: Among them, ϵ is set to an arbitrary small value to prevent division by zero. * Finally, the global encoding features are guided by introducing this preference of the source domain into the shared domain Thus effectively extracting the final user features in the target domain The linear layer g u With parameter φ, it is used to dynamically adapt the encoded features to the changing preference features and extract the final user features: Objective function of prototype learning The definition is as follows: Among them, c is a fully connected layer used to predict the probability of positive or negative, g p is the prototype layer, represents the cross entropy loss for preference classification within the prototype branch; l is an array used to distinguish item categories, and λ1, λ2, and λ3 are weights for controlling the loss; To build the final prototype, the constraints considered included: (1) Setting clustering loss To encourage these interaction items in the source domain to approach a prototype corresponding to their preferences; (2) Setting separation loss to promote an increase in the distance between the characteristics of the coded item and the prototype that does not belong to its preference; (3) Setting diversity loss By penalizing prototypes that are too close, the diversity of the learned prototypes is encouraged; In the above formula, yes The prototype set under the class, k = {1, 2} represents two preferences, ξ is the threshold of cosine similarity in diversity loss; Finally, the user's corresponding project prediction score is passed Calculation to determine recommended projects; This method also uses binary cross entropy loss to predict the loss: in, Indicates uninteresting items by dot product calculation Negative prediction score.

Images