Dynamic cross-domain recommendation method for target perception complementary knowledge migration and information fusion

Abstract

The invention relates to a dynamic cross-domain recommendation method for target perception complementary knowledge migration and information fusion. The method comprises a target perception personalized bridge module used for extracting personalized preference characteristics of a user, a public bridge module used for migrating public preference characteristics of the user and a dynamic target perception fusion module. Firstly, a target perceptual bridge module is proposed to integrate target information to guide a learning process of a personalized bridge; secondly, learning public preferences of a public bridge modeling user; and finally, providing a dynamic target perception fusion module to carry out complementary fusion on the output of the personalized bridge and the public bridge. The target commodity information is considered in the fusion process, so that the fusion process is dynamically changed along with the change of the target commodity characteristics, and the model performance is further improved; and finally, experiments on three different cross-domain tasks all obtain optimal performance, and the effectiveness of the model is verified.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross - domain recommendation, and relates to a dynamic cross - domain recommendation method for target - aware complementary knowledge transfer and information fusion. Background Art

[0002] With the advent of the information age, the problem of information overload has become increasingly prominent. Recommendation systems have received more and more attention because they can alleviate the problem of information overload to a certain extent and are widely used in various occasions, such as e - commerce systems, social network systems, etc. At the same time, many excellent recommendation models have emerged. However, these works mainly focus on how to model the historical interaction information of users and the attribute features in other fields to characterize user preferences and make recommendations based on these preferences. Although quite good results have been achieved, in real - world scenarios, for new users, since they do not have historical interaction information and lack attribute features in this field, it is difficult for such recommendation systems to accurately characterize the preferences of new users, so accurate recommendations cannot be made for new users.

[0003] As a very promising solution to this cold - start problem, cross - domain recommendation technology has gradually received more and more attention from researchers in academia and industry in recent years. Its main idea is that the preferences of the same user between the source domain and the target domain are considered to be relevant, and the cold - start problem in the target domain can be alleviated by migrating the knowledge in the source domain with relatively rich user information to the target domain. Therefore, for cross - domain recommendation systems, how to better transfer the knowledge in the source domain to the target domain has become the key issue. Some existing works, based on the assumption of "relationship sharing between all users in the source domain and the target domain", learn a common mapping function for all users between the source domain and the target domain to achieve the transfer of source - domain knowledge and alleviate the cold - start problem in the target domain.

[0004] However, this assumption is not practical. The relationships between the preferences of different users in the source domain and the target domain may not be exactly the same, resulting in the difficulty of accurately reflecting the preference information of each user by the method of learning a common mapping function between the source domain and the target domain for knowledge transfer. For this problem, the existing PTUPCDR model proposes a solution of providing personalized mapping functions for each user, further considering the different preference information of different users, and better overcoming the drawbacks of the method based on a common mapping function, thus improving the performance of cross - domain recommendation systems.

[0005] However, both the method based on a common mapping function and the method of assigning a personalized mapping function to each user have their inevitable limitations. Either it cannot accurately reflect user - personalized preferences, resulting in inaccurate recommendations; or it only focuses on user - personalized preferences, easily leading to information cocoons and reducing the diversity of recommendation results.

[0006] At the same time, the existing methods of personalized mapping functions do not take into account the personality characteristics of the target commodity, which may lead to the interference of noise information irrelevant to the personality characteristics of the target commodity, and further lead to the decline of recommendation performance. Summary of the Invention

[0007] In view of this, in order to solve the above problems, the present invention provides a dynamic cross-domain recommendation method for target-aware complementary knowledge transfer and information fusion, proposes a target-aware personalized bridge module to integrate target information to guide the learning process of the personalized bridge, learns a common bridge to model the common preferences of users, and proposes a dynamic target-aware fusion module to complementarily fuse the outputs of the personalized bridge and the common bridge, taking into account the target commodity information during the fusion process, and improving the model performance.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A dynamic cross-domain recommendation method for target-aware complementary knowledge transfer and information fusion, the dynamic cross-domain recommendation is based on a cross-domain recommendation model DnmCDR, and includes a target-aware personalized bridge module for extracting user personalized preference features, a common bridge module for migrating user common preference features, and a dynamic target-aware fusion module;

[0010] The target-aware personalized bridge module introduces the target commodity feature information while considering the user's personalized preferences, can avoid the interference of noise information irrelevant to the target commodity, and the obtained user representation can more accurately reflect the preference degree for the target commodity, so as to make a more accurate scoring judgment;

[0011] The common bridge module shares a mapping function for all users, so it can consider the common preference relationship between users in two domains, which can avoid possible information cocoon problems to a certain extent and improve the diversity of recommendation results;

[0012] The dynamic target-aware fusion module, while fusing the user's individual preferences and common preferences, takes into account the target commodity feature information for the selection and fusion of the user's individual and common information again. On the one hand, it can further avoid the interference of information irrelevant to the target commodity in the user representation, and on the other hand, it can balance the individual and common preferences and further improve the model performance;

[0013] In this dynamic cross-domain recommendation, it includes two domains, a source domain and a target domain, both of which include their respective corresponding user sets U = {u1, u2,..., u |U|}, item sets V = {v1, v2,..., v |V|}, and a rating set R composed of user ratings for items, r ij$r_{ij} \in \mathbb{R}$ represents the rating of user $i$ on item $j$. For each user in the source domain Let represent the corresponding item interaction sequence in the source domain, where $t$ i represents the number of items interacted with.

[0014] Furthermore, the target-aware personalized bridge module is specifically as follows: The dual-channel attention module is used to model both the item sequence interacted by the user in the source domain and the target item simultaneously, and to depict the preference features of both the user and the target item, obtaining the source-domain user preference representation considering the preference features of the target item; Based on the preference representation, the personalized mapping module is used to map this user representation from the source domain to the target domain, obtaining the corresponding user representation in the target domain. Beneficial effect: The user representation in the target domain considers the preference features of the target item, avoiding the interference of information unrelated to the target item in the user representation.

[0015] Furthermore, the dual-channel attention module includes two feature encoders, Attn src (·) and Attn tgt-item (·). For the current user its preference feature representation is obtained by weighted fusion of the representations of the items it has interacted with:

[0016]

[0017] where Attn src (·) represents the corresponding attention network module, and $\theta$ src represents its corresponding model parameters, represents the $l$-th item interacted by user in the source domain.

[0018] Beneficial effect: The attention weight scores for different items calculated according to the dual-channel attention module can reflect the importance degree of different items for learning the personalized bridge.

[0019] Furthermore, the concatenation of the embedded representations of the items interacted by the current user in the source domain and the embedded representation of the target item is used as the input to the attention module to calculate the corresponding attention weights for weighted fusion, obtaining the preference representation of the current user, which is specifically:

[0020]

[0021] where Attn tgt-item (·) represents the corresponding attention network module, and $\theta$ tgt-item represents its corresponding model parameters, represents user The l-th interacted commodity in the source domain represents the target commodity.

[0022] Furthermore, a personalized mapping module is adopted to model the personalized preference features of users and commodities. Specifically: For and they are concatenated as the input of a meta-network, and the parameters of the personalized mapping module are generated by this meta-network. The source domain user representation fused with the target commodity information is mapped to the target domain by using the personalized mapping module. The specific process is as follows:

[0023]

[0024] where, g(·) represents the meta-network, represents the learnable parameters corresponding to the meta-network;

[0025] is the output vector of the meta-network, and its size depends on the corresponding mapping function structure. The shape of is deformed into d×d to obtain the parameters of the personalized mapping module,

[0026]

[0027] Taking as the parameters of the personalized mapping module, we get the representation in the target domain

[0028]

[0029] Beneficial effects: The target domain user representation obtained through the personalized mapping module takes into account both the personalized preference features of the user and the preference features of different commodities.

[0030] Furthermore, in the common bridge module, all users share a common mapping function, which focuses on the common features existing in the mapping of all users in the source domain to the target users. Specifically: The common mapping function directly takes the source domain user representation as the input, and all users share the common bridge parameters, and the output is the representation of this user in the target domain after the source domain user passes through the common bridge mapping and fuses the common preference features of the user. The specific formalization is as follows:

[0031]

[0032] where, f c (·) adopts a two-layer perceptron structure, $\mathbf{e}_{i}$ is the embedded representation of user $i$ in the source domain, and $\omega$ is the learnable parameter. Beneficial effects: It can improve the diversity of recommendation results to a certain extent, avoid possible information cocoon problems, and can still provide a good mapping result when the interaction of source domain users is less, resulting in inaccurate source domain user representations.

[0033] Furthermore, the dynamic target-aware fusion module can model the features that change with the change of users, especially target commodities. The specific process is as follows:

[0034] S1: Obtain the corresponding Q, K, and V in the attention mechanism, and the representation in the source domain are respectively and concatenated to obtain and

[0035] S2: Then stack and to obtain and use as K and V in the attention mechanism;

[0036] S3: The target commodity representation $\mathbf{v}$ t is respectively and concatenated and then stacked to obtain which is used as Q in the attention mechanism.

[0037] Furthermore, the specific process of the dynamic target-aware fusion module is as follows:

[0038]

[0039] Perform attention fusion output to obtain

[0040]

[0041] where $sm(\cdot)$ represents the softmax function, and are learnable parameters;

[0042] Split row-wise into two d-dimensional vectors, denoted by and respectively, where is the output corresponding to after passing through the dynamic target-aware fusion module; where is

[0043] Furthermore, a dynamic target-aware fusion module is adopted to explicitly model the above-mentioned features that vary with the user and the target commodity, obtaining a representation that further fuses the features of the source-domain user and the target commodity. and

[0044] The and are mean-fused to obtain the final user representation.

[0045]

[0046] The score of the user for the target commodity is calculated through the inner product of and :

[0047]

[0048] The final objective function is:

[0049]

[0050] where is the set of true scores of common users in the target domain, is the predicted score of the model.

[0051] Furthermore, the model DnmCDR is a mapping-based cross-domain recommendation model.

[0052] The beneficial effects of the present invention are as follows:

[0053] 1. The dynamic cross-domain recommendation method for target-aware complementary knowledge transfer and information fusion disclosed by the present invention adopts a dynamic cross-domain recommendation model for target-aware complementary knowledge transfer and information fusion. First, aiming at the problem that the existing method of personalized mapping function fails to consider the individual characteristics of the target commodity, a target-aware personalized bridge is proposed, taking into account the individual preference characteristics of both the user and the target commodity. Second, aiming at the problem that only focusing on the personalized preferences of users is likely to lead to information cocoons and reduce the diversity of recommendation results, a common bridge is proposed, which can alleviate the possible information cocoon problem to a certain extent and improve the diversity of recommendation results. Finally, aiming at the problem that the importance of the individual and common characteristics of users for target commodity recommendation may change with the changes of users and target commodities, a dynamic target-aware fusion module is proposed to model the characteristics that change with the changes of users and target commodities. This recommendation method has achieved optimal performance in three different cross-domain recommendation tasks, proving its effectiveness.

[0054] 2. The dynamic cross - domain recommendation method of target - aware complementary knowledge transfer and information fusion disclosed by the present invention proposes a target - aware personalized bridge, which takes into account the personalized preference characteristics of both users and target products, and solves the problem that the existing method of personalized mapping function does not consider the personalized characteristics of target products.

[0055] 3. The dynamic cross - domain recommendation method of target - aware complementary knowledge transfer and information fusion disclosed by the present invention proposes a common bridge, which can alleviate the possible information cocoon problem to a certain extent, improve the diversity of recommendation results, and solve the problem that only focusing on the personalized preferences of users is likely to lead to the information cocoon and reduce the diversity of recommendation results.

[0056] 4. The dynamic cross - domain recommendation method of target - aware complementary knowledge transfer and information fusion disclosed by the present invention proposes a dynamic target - aware fusion module to model the characteristics whose importance changes with the changes of users and target products, and solves the problem that the importance of the personalized and common characteristics of users for target product recommendation may change with the changes of users and target products.

[0057] Other advantages, objectives and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0059] Figure 1 is the overall framework diagram of the model DnmCDR in the dynamic cross - domain recommendation method of target - aware complementary knowledge transfer and information fusion of the present invention;

[0060] Figure 2 (a) is the performance comparison diagram of the model DnmCDR and the baseline models EMCDR and PTUPCDR on the MF, GMF and YouTube DNN base models respectively when the proportion of the test set is 20% in this embodiment; Figure 2 (b) is the performance comparison diagram of the model DnmCDR and the baseline models EMCDR and PTUPCDR on the MF, GMF and YouTube DNN base models respectively when the proportion of the test set is 50% in this embodiment; Figure 2 (c) is the performance comparison diagram of the model DnmCDR and the baseline models EMCDR and PTUPCDR on the MF, GMF and YouTube DNN base models respectively when the proportion of the test set is 80% in this embodiment. Detailed implementation manners

[0061] The following describes the implementation manners of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. The related problems in the prior art are as follows:

[0062] I. Cold start problem; the cold start problem is a very critical problem in the recommendation system. In today's era of information explosion, people are constantly facing the impact of a vast amount of information. However, for individuals, it is undoubtedly very difficult, or even impossible, to select the appropriate information or products they like from the vast amount of information. Therefore, it becomes very important to propose a method to help people select appropriate information or products in the environment of information overload. In this context, the recommendation system emerged. It can better solve the problem of information overload by understanding the main preferences of users, and recommend to users the parts of information or products that they may be interested in from the vast amount of information or products. However, for cold start users, since the recommendation system does not know their preference characteristics, it is impossible to recommend appropriate information or products to users. This is the so-called cold start problem. It is very challenging to recommend appropriate information or products to cold start users because the lack of necessary user-product interaction data makes it difficult for existing recommendation systems to accurately characterize the characteristics of users and products, resulting in unsatisfactory recommendation performance. However, in today's era of information explosion, new users and new products are constantly emerging, and the cold start problem is even more normal, accompanying the entire cycle of the recommendation system; in addition, for enterprises, in order to achieve greater economic benefits, their recommendation systems have to develop in the direction of attracting more new users and enhancing user stickiness, which in turn forces the recommendation system to have better cold start recommendation performance. Therefore, how to solve the cold start problem well has become the core of the optimization of the recommendation system.

[0063] Existing work mainly uses some auxiliary information in the field to alleviate the cold start problem. For example, for new users, prior knowledge can be used to recommend popular products to new users; it is also possible to make recommendations by using the basic information filled in by users during registration, such as gender, age, hobbies, industry, etc.; for new products, the information of the products themselves can be used for recommendation, such as category, use, price and other descriptive information; in addition, in recent years, cross-domain recommendation systems have achieved very good results in alleviating the cold start problem in the target domain by making full use of the interaction information between users and products in the source domain, attracting more and more researchers to study it, and various models have been derived.

[0064] II. Cross - domain recommendation; For a long time, how to provide accurate recommended content for a new user without interaction history has been an inevitable and very challenging problem. In a recommendation system, such a user is called a cold - start user, and such a problem is called a cold - start problem. And such cold - start problems often become the bottleneck restricting the performance of the recommendation system. Therefore, in order to improve the performance of the recommendation system, it is very necessary to alleviate such cold - start problems.

[0065] In recent years, the research on cold - start problems in recommendation systems has attracted more and more researchers, and various models have been continuously derived. Among them, a simple but very effective method for solving cold - start problems has attracted more and more attention from researchers. It helps the target domain alleviate the cold - start problem by making full use of the interaction information between users and items in the source domain to improve the recommendation performance of the target domain. A recommendation system that alleviates the cold - start problem of the target domain according to this mode is called a cross - domain recommendation system.

[0066] In recent years, many excellent cross - domain recommendation system models have emerged one after another, continuously improving the performance of the recommendation system, further demonstrating the effectiveness of the source - domain interaction information of users in solving the cold - start problem of the target domain and improving the performance of the recommendation system. Existing scholars have proposed the CMF model based on the matrix factorization model (MF). In the CMF model, the overlapping users in the source domain and the target domain share the embedding representation, which is an extension of the matrix factorization model; the proposed CST model initializes the corresponding user representation in the target domain by using the pre - trained user vector representation in the source domain, and at the same time constrains the common users and items to make their vector representations in the two domains as close as possible; the proposed CoNet model realizes two - way knowledge transfer by interacting knowledge in the two domains to achieve two - way cross - domain recommendation; the proposed MINDTL model first extracts the rating pattern information of source - domain users from the source - domain user - item rating set, and then combines the word - rating pattern information with the target - domain information to make recommendations for the target domain; based on the assumption that both the source domain and the target domain have their own domain knowledge, the DDTCDR model is proposed. This model extracts the knowledge of users in the source domain and the target domain while retaining the relationship of the latent space representation of users in different domains; the proposed EMCDR model based on embedding and mapping first learns the user and item embedding representations of the source domain and the target domain through single - domain models respectively, then learns a common mapping function on the basis of the learned embedding representations, and then transfers the preferences of users in the source domain to the target domain for recommendation; the PTUPCDR model based on meta - learning is proposed. This model is an improvement of the EMCDR model, which uses the meta - learning technology based on parameter generation to learn a preference transfer function for each user, enabling the model to fully consider the individual preferences of each user; the proposed DiffCDR model regards the process of restoring the samples with added noise as a data transformation process, thus establishing a connection between the diffusion probability models (DPMs) and the mapping module in cross - domain recommendation and replacing the mapping module with this. Although the above methods have achieved good results, on the one hand, these methods all ignore the guiding role of target - item information in the model training process; on the other hand, these methods either learn a common bridge or a personalized bridge, but ignore the potential complementarity between the two. Therefore, based on the above observations, this paper proposes the DnmCDR model to solve the above limitations.

[0067] III. Meta - learning; Meta - learning is a "learning to learn" paradigm, aiming to quickly adapt to new tasks by training on the distribution of similar tasks. Recently, meta - learning methods have achieved great success in the fields of computer vision and natural language processing. Inspired by this, various meta - learning methods have been proposed to alleviate the cold - start problem. These methods can be divided into three different research directions, including parameter - generation - based methods, gradient - based methods, and metric - based methods. Existing scholars have proposed the MeLU model to infer the preferences of new users through a small number of consumed items and introduced an evidence candidate selection strategy to determine the discriminative items for customized preference inference. There are also those who use parameter - generation - based meta - learning techniques to learn a preference transfer function for each user, enabling the model to fully consider the personalized preferences of each user. By learning a task - oriented meta - network in the meta - stage, the user preferences are implicitly transferred from the source domain to the target domain. The DnmCDR proposed in this paper belongs to the parameter - generation - based method, which generates parameters by simultaneously considering the interactive item features in the source domain and the target item features.

[0068] As Figure 1 shown, a dynamic cross - domain recommendation method for target - aware complementary knowledge transfer and information fusion, the dynamic cross - domain recommendation is based on the cross - domain recommendation model DnmCDR. The model DnmCDR is a mapping - based cross - domain recommendation model, including three modules: target - aware personalization bridge, common bridge, and dynamic target - aware fusion.

[0069] In a cross - domain recommendation system, there are usually two domains: source and target. Among them, both the source domain and the target domain contain their respective user sets U = {u1, u2,..., u |U|}, item sets V = {v1, v2,..., v |V|}, and rating sets R composed of user ratings for items. Here, |U| and |V| represent the number of users and items in the current domain, and r ij ∈R represents the rating of user i for item j. To avoid confusion in the definition of the source domain and the target domain, this paper uses the superscript symbol l∈{s, t} for distinction. Specifically, U s , V s and R s are used to represent the user, item sets, and their corresponding rating sets in the source domain respectively. Correspondingly, U t , V t and R t are used to represent the user, item sets, and their corresponding rating sets in the target domain respectively. Let U o = U s ∩U t represent the common user set of the source domain and the target domain, Denote the set of cold start users. Note that in a cross - domain recommendation system, there are no common items between the source domain and the target domain, that is

[0070] For each user in the source domain Let denote the corresponding item interaction sequence in the source domain, where t i denotes the number of interacted items.

[0071] In the single - domain latent factor model, each user and item are assigned a low - dimensional embedding representation, also known as an embedding vector. In this paper, use to represent user embeddings and item embeddings, where d represents the assigned embedding dimension, and l ∈ {s, t} is used to distinguish the embedding representations of the source domain and the target domain.

[0072] For the target - aware personalized bridge module, it is mainly used to extract the personalized preference features of users. This module introduces the target item feature information while considering the user's personalized preferences. Therefore, the target - domain user representation obtained through this module can avoid the interference of noise information unrelated to the target item, and thus the obtained user representation can more accurately reflect the preference degree for the target item, and then make a more accurate scoring judgment.

[0073] According to the attention mechanism, for the current user For a certain item interacted with, the attention score calculated by the attention module for this item can reflect its importance for the current user's personalized bridge function. That is to say, the current user's personalized preferences can be characterized by the attention - weighted fusion of the items interacted with. However, the current work does not consider the preference information of the target item when learning the personalized bridge function, which may lead to sub - optimal final recommendation results because the correlation between the personalized bridge - transferred user preference features and the target item features is not direct and may introduce noise, resulting in a decline in the performance of the model. Therefore, to solve this problem, this paper first proposes a dual - channel attention module to simultaneously model the item sequence interacted with by the user in the source domain and the target item, and then can simultaneously characterize the preference features of the user and the target item, obtaining a source - domain user preference representation considering the target item preference features; then, based on this representation, this paper further proposes a personalized mapping module to map this user representation from the source domain to the target domain, obtaining the corresponding user representation in the target domain. In this way, this target - domain user representation also considers the target item preference features and avoids the interference of information unrelated to the target item in the user representation.

[0074] Among them, the dual - channel attention module mainly includes two feature encoders based on the attention mechanism, namely Attn src (·) and Attn tgt-item(·) For a personalized bridge, the attention weight scores for different products calculated according to the attention module can reflect the importance degree of different products for learning the personalized bridge. That is to say, for the previous user its preference feature representation can be obtained by weighted fusion of the product representations it has interacted with (i.e., Attn src (·)):

[0075]

[0076]

[0077] where Attn src (·) represents the corresponding attention network module, θ src represents its corresponding model parameters, represents the user the l-th interacted product of the user in the source domain.

[0078] As mentioned above, the output of Attn src (·) does not take into account the target product feature information, which may lead to sub-optimal recommendation results. Therefore, this paper uses another attention module, namely Attn tgt-item (·) to model the target product feature information. Specifically, first, the product embedding representations of the current user interacted in the source domain and the target product embedding representation are concatenated, and then the concatenated representation sequence is used as the input of the attention module to calculate the corresponding attention weights for weighted fusion, and finally the preference representation of the current user is obtained. Obviously, this user representation takes into account the preference information of the target product. The specific process is as follows:

[0079]

[0080] where Attn tgt-item (·) represents the corresponding attention network module, θ tgt-item represents its corresponding model parameters, represents the user the l-th interacted product of the user in the source domain, represents the target product.

[0081] For the source domain and the target domain, the preference relationships of different users between the two domains are not exactly the same; similarly, the preference features corresponding to commodities also change with the change of commodities. That is to say, the personalized mapping function should take into account both the personalized preference features of different users and the preference features of different commodities. Considering that the personalized mapping method based on meta-learning can take into account the personalized preferences of users, this paper proposes a similar personalized mapping module to model the personalized preference features of users and commodities. Specifically, first, and are concatenated as the input of a meta-network, and then the parameters of the personalized mapping module are generated by this meta-network. Finally, the source domain user representation fused with the target commodity information is mapped to the target domain by the personalized mapping module. In this way, the target domain user representation obtained by the personalized mapping module takes into account both the personalized preference features of users and the preference features of different commodities. The specific process is as follows:

[0082]

[0083] Among them, g(·) represents the meta-network, represents the learnable parameters corresponding to the meta-network;

[0084] is the output vector of the meta-network, and its size depends on the corresponding mapping function structure. The shape of is deformed into d×d to obtain the parameters of the personalized mapping module,

[0085]

[0086] Taking as the parameters of the personalized mapping module, we get the representation in the target domain

[0087]

[0088] As mentioned above, for the target-aware personalized bridge module, it has comprehensively considered the personalized preference characteristics of users and target products. However, in a recommendation system, overemphasis on the personality information of products or users may instead fall into the situation of an information cocoon, resulting in suboptimal results because users' preferences are not static and may be dynamic and diverse. Therefore, this paper introduces a common bridge module. In this module, all users share a common mapping function, which focuses on the common characteristics existing in the mapping from all users in the source domain to the target user. This can not only improve the diversity of recommendation results to a certain extent and avoid possible information cocoon problems, but also provide a better mapping result when there is less interaction among source domain users, resulting in inaccurate representation of source domain users (because the common bridge takes into account the common preferences of most users). Specifically, the common mapping function directly takes the source domain user representation as input, and all users share the common bridge parameters, and the output is the representation of this user in the target domain after being mapped by the common bridge and integrating the common preference characteristics of the user. It can be formalized as follows:

[0089]

[0090] where f c (·) adopts a two-layer perceptron structure, is the embedded representation of user i in the source domain, and ω is a learnable parameter.

[0091] So far, the representation of the user in the target domain considering both the personalized preference characteristics of the user and the target product personalized preference characteristics and the representation in the target domain considering the common characteristics of all users have been obtained. However, the importance of these two user representations for target product recommendation may change with the changes of users and target products. Therefore, in order to further improve the model performance, it is necessary to explicitly model the characteristics that change with the changes of users and target products. Although the attention mechanism does have the ability to separately model the importance of these two representations and fuse them according to their importance, the importance of these representations relative to the target item is static. In other words, they cannot dynamically reflect the characteristics of the target item, affecting the accuracy of fusing user representations, and thus resulting in inaccurate scoring of the target item. Inspired by the CAUM and MiceRec methods, this paper proposes a dynamic target-aware fusion module, which can model the characteristics that change with the changes of users, especially target products. The specific process is as follows:

[0092] S1: Obtain the corresponding Q, K, and V in the attention mechanism, and take the representation in the source domain respectively with and concatenated to obtain and

[0093] S2: Then stack and to obtain Take as K and V in the attention mechanism;

[0094] S3: The target commodity representation v t respectively with and After concatenation and then stacking, obtain Take it as Q in the attention mechanism.

[0095] The specific process is as follows:

[0096]

[0097] Perform attention fusion output to obtain

[0098]

[0099] where sm(·) represents the softmax function, and are learnable parameters;

[0100] For ease of description, this article splits by row into two d-dimensional vectors, respectively represented by and where is the output corresponding to after passing through the dynamic target perception fusion module; where is the output corresponding to after passing through the dynamic target perception fusion module.

[0101] Through the target perception personalized bridge and the public bridge, the representations of the source domain users in the target domain after being transformed by these two bridges are obtained, which are respectively and However, the importance of these two user representations for target commodity recommendation changes with the change of users, especially the target commodity. Therefore, this article proposes a dynamic target perception fusion module to explicitly model the features that change with users and target commodities, and better solves this problem, obtaining a representation that further fuses the features of source domain users and target commodities and

[0102] Take and Perform mean fusion to obtain the final user representation

[0103]

[0104] Through and Inner product to calculate the user Rating for the target item :

[0105]

[0106] The final objective function is:

[0107]

[0108] Where The set of true ratings of common users in the target domain, Is the predicted rating of the model

[0109] The following is illustrated by specific experiments:

[0110] I. Experimental Setup

[0111] 1) Dataset establishment

[0112] To verify the effectiveness of the proposed model in real scenarios, similar to many recommendation models, this paper selects the Amazon dataset as the evaluation dataset. The Amazon dataset is a large-scale public dataset recording user-item interaction behaviors in real scenarios, which contains a total of 24 different item domains. This paper selects three of them for experiments, namely movies_and_tv (Movie), cds_and_viny1 (Music), and books (Book). And three different cross-domain recommendation tasks are defined according to these three domains, namely Task 1: Movie - Music, Task 2: Book - Movie, and Task 3: Book - Music. For each cross-domain recommendation task, a part of the common users is randomly sampled as the test set, and the remaining part is used as the training set. Table 1 shows the detailed statistical information of the three different cross-domain tasks

[0113] Table 1 Statistical Information of Different Cross-Domain Task Data

[0114]

[0115] 2) Evaluation Metrics

[0116] Essentially, cross-domain rating prediction belongs to a regression problem. Therefore, similar to many recommendation models

[0117] Consistent with this, the mean absolute error (MAE) and root mean square error (RMSE) used in this paper are used as model performance evaluation indicators. For the convenience of explanation, the mean absolute error and root mean square error are represented by MAE and RMSE respectively in the following text. Obviously, whether it is MAE or RMSE, the smaller the value, the better the model performance. Let n represent the number of samples, and y i respectively represent the predicted score and the true score of the model corresponding to sample i. Then MAE and RMSE can be defined as follows:

[0118] MAE: It measures the average of the absolute errors between the predicted values and the true values.

[0119]

[0120] RMSE: It measures the deviation between the predicted values and the true values.

[0121]

[0122] 3) Baseline models

[0123] The cross-domain recommendation model proposed in this paper belongs to the mapping-based cross-domain recommendation algorithm. Therefore, in order to fairly evaluate the effectiveness of the method proposed in this paper, the following mapping-based cross-domain recommendation algorithms are selected for comparison:

[0124] TGT. TGT represents the target domain matrix factorization model, which is trained only using the target domain data.

[0125] CMF. CMF can be regarded as an extension of the matrix factorization model for cross-domain recommendation problems. In this model, the source domain and target domain users share the embedding representation.

[0126] EMCDR. EMCDR first proposed using a mapping-based framework to solve cross-domain recommendation problems. It first projects users and items into their respective latent spaces in the source domain and target domain respectively, and then learns a mapping function between the latent spaces of the two domains to model the relationship between the two domains.

[0127] DCDCSR. DCDCSR first considered the impact of the sparsity of user-item interactions on the accuracy of the representations learned in the space. It uses a fully connected neural network as the mapping function and models the sparsity of user-item interactions into the mapping function.

[0128] SSCDR. SSCDR considered the problem that it is difficult to learn a good mapping function when there are too few overlapping users between the source domain and the target domain, and proposed a semi-supervised method, so that a good mapping function can be learned even when there are only a small number of overlapping users between the two domains.

[0129] PTUPCDR. PTUPCDR explicitly models the problem of personalized mapping relationships between different users in two domains for the first time using meta-learning techniques.

[0130] DiffCDR. DiffCDR replaces the mapping module with DPMs in CDRs by regarding the recovery of images from noisy samples as a data transformation process. Specifically, DiffCDR designs a diffusion module based on DPM theory to generate user embeddings in the target domain and also considers the distribution of the target embedding space. In addition, DiffCDR adopts an alignment module and a task-oriented loss function to reduce the negative impact brought by the randomness introduced in the diffusion module. Therefore, DiffCDR is more stable and can adapt to specific tasks. It is the optimal baseline comparison model in this paper.

[0131] II. Overall Experiments

[0132] The overall comparison results of the proposed DnmCDR model in this paper and all baseline models in terms of the MAE and RMSE evaluation metrics are shown in Tables 2 and 3. The experimental results show that the proposed model in this paper has achieved the optimal performance in all tasks and has statistical significance. The performance of all baseline comparison models in this paper adopts the results in the original paper. In order to verify that the optimal result is statistically significant compared with the sub-optimal result, the results of running randomly five times are used for statistical analysis:

[0133] The DnmCDR model in this paper has significantly improved the recommendation performance in all tasks and has a significant improvement in statistical significance. For example, taking the MAE evaluation metric as an example, for Task 1, when the test set ratio is 20%, DnmCDR has improved by 24.9% compared with the optimal baseline model; regardless of any one of the three cross-domain recommendation tasks, when the test set ratio is smaller (the training set ratio is larger), the performance improvement of DnmCDR compared with the optimal baseline model is also smaller. For example, for Task 1, when the test set ratios are 20%, 50%, and 80% respectively, the performance improvements of DnmCDR compared with the optimal baseline model are 24.9%, 33.0%, and 40.8% respectively; compared with Task 1 and Task 3, the improvements of DnmCDR compared with the optimal baseline model in Task 2 are relatively smaller (but also significant), which are 9.3%, 12.4%, and 16.1% respectively. Combining the statistical analysis of the dataset in Table 1, it is found that the number of overlapping users in the two domains in Task 2 is approximately twice that of Task 1 and Task 3. The reason is that more overlapping users can provide richer source domain information, and thus can learn better mapping functions by themselves, making the improvement space of the model in this paper relatively smaller.

[0134] Table 2 Comparison of the proposed model (Ours) with other baseline models in terms of the mean absolute error (MAE↓) evaluation metric

[0135]

[0136] Improve% represents the relative percentage improvement in the performance of the proposed model compared to the optimal baseline model (such as DiffCDR). The optimal results are in bold. Additionally, * indicates that through the T - test, p ≤ 0.05, the proposed model has a statistically significant improvement in performance compared to the optimal baseline model.

[0137] Table 3 Comparison of the proposed model (Ours) with other baseline models in terms of the root mean square error (RMSE↓) evaluation metric

[0138]

[0139] Improve% represents the relative percentage improvement in the performance of the proposed model compared to the optimal baseline model (such as DiffCDR). The optimal results are in bold. Additionally, * indicates that through the T - test, p ≤ 0.05, the proposed model has a statistically significant improvement in performance compared to the optimal baseline model.

[0140] III. Ablation Experiments

[0141] To reveal the contribution of each sub - module proposed in this paper to the overall model, ablation experiments were performed as shown in Tables 4 and 5. The ablation details of each module are explained as follows:

[0142] w / o tgt - per: Removing the target - aware personalized bridge and only using the output of the common bridge mapping as the final target - domain user representation.

[0143] w / o com: Removing the common bridge part and only using the target - aware personalized bridge for mapping, and taking its output as the final target - domain user representation.

[0144] w / o tgt - fus: Removing the target - aware fusion module and directly using the mean - pooling result of the outputs of the target - aware personalized bridge and the common bridge as the final target - domain user representation.

[0145] Tables 4 and 5 respectively show the ablation experiments of the proposed model DnmCDR on each sub-module. According to the tables, on all tasks, each sub-module of DnmCDR has an effect, and the performance decreases after removing each sub-module; among them, the performance decreases most significantly when removing the target-aware personalized bridge, decreases relatively significantly when removing the common bridge module, and decreases relatively insignificantly when removing the target-aware fusion module;. For example, for Task 1, when the proportion of the training set is 80% (β = 20%), when removing the target-aware personalized bridge, the MAE and RMSE indicators increase by 0.2333 and 0.3011 respectively, when removing the common bridge, the MAE and RMSE indicators increase by 0.1723 and 0.2147 respectively, and when removing the target-aware fusion module, the MAE and RMSE indicators increase by 0.0181 and 0.0463 respectively. (Note that for the MAE and RMSE indicators, the smaller the value, the better the performance). This shows that compared with the target commodity features in the fusion stage, the personalized and common features of target awareness of source domain users in the mapping stage are more important for improving the model performance.

[0146] Table 4 Ablation Experiment for Mean Absolute Error (MAE)

[0147]

[0148] Table 5 Ablation Experiment for Root Mean Square Error (RMSE)

[0149]

[0150] IV. Generalization Experiment

[0151] Some current cross-domain recommendation methods, including the method proposed in this paper, all use the matrix factorization (MF) model as the base model for single-domain training to perform experimental evaluations of the performance of their respective cross-domain recommendation models. To verify the performance of the model in this paper on different base models, this paper further evaluates the performance of DnmCDR on different base models and compares it with other cross-domain models. Refer to Figure 2 , which shows the performance of the model in this paper and the other two baseline models PTUPCDR and EMCDR on MF, GMF, and YouTube DNN respectively. Figure 2 (a) is a performance comparison diagram of the model DnmCDR and the baseline models EMCDR and PTUPCDR on the MF, GMF, and YouTube DNN base models when the proportion of the test set is 20%; Figure 2 (b) is a performance comparison diagram of the model DnmCDR and the baseline models EMCDR and PTUPCDR on the MF, GMF, and YouTube DNN base models when the proportion of the test set is 50%; Figure 2(c) Performance comparison diagrams of the DnmCDR model with the baseline models EMCDR and PTUPCDR on the MF, GMF, and YouTube DNN base models when the proportion of the test set is 80%.

[0152] The following conclusions can be drawn from the experimental results: Regardless of whether MF, GMF, or YouTube DNN is used as the base model, DnmCDR achieves the best performance compared to EMCDR and PTUPCDR; regardless of whether it is the EMCDR, PTUPCDR, or the DnmCDR model in this paper, overall, the performance with MF as the base model is the worst, the performance with YouTube DNN as the base model is the best, and the performance with GMF as the base model is between the two; conversely, DnmCDR has the largest improvement compared to the PTUPCDR model when MF is used as the base model, the smallest improvement compared to the PTUPCDR model when YouTube DNN is used as the base model, and the improvement compared to the PTUPCDR model when GMF is used as the base model is between that with MF and YouTube DNN as the base models.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A dynamic cross-domain recommendation method based on target-aware complementary knowledge transfer and information fusion, characterized in that: The dynamic cross-domain recommendation is based on a cross-domain recommendation model DnmCDR, including a target-aware personalized bridge module for extracting user personalized preference features, a public bridge module for migrating user public preference features, and a dynamic target-aware fusion module; The target-aware personalized bridge module considers the personalized preferences of users and introduces the characteristic information of target products, which can avoid the interference of noise information irrelevant to the target products. The obtained user representation can accurately reflect the preference for the target products and make accurate scoring judgments. In the public bridge module, all users share a mapping function, so the common preference relationship between users in the two domains can be taken into account, the information cocoon problem can be avoided, and the diversity of recommendation results can be improved; The dynamic target perception fusion module, while integrating the user's individual preferences and common preferences, again considers the selection and fusion of the user's individual and common information by the target product feature information, on the one hand avoiding the interference of irrelevant information about the target product in the user's representation, on the other hand balancing the individual and common preferences to improve the model performance; In the dynamic cross-domain recommendation, there are two domains, source and target, each of which includes a corresponding user set U = {u1, u2, ..., u |U| }、Product set V={v1,v2,...,v |V| } and the rating set R, r consisting of the user's rating of the product ij ∈R represents the rating of user i on product j. For each user in the source domain make represents the commodity interaction sequence in the corresponding source domain, where t i Indicates the number of items that have been interacted with.

2. The target-aware complementary knowledge transfer and information fusion dynamic cross-domain recommendation method according to claim 1, characterized in that: The target perception personalized bridge module is specifically: using a dual-channel attention module to simultaneously The interacted product sequence and target product are modeled, and the preference characteristics of users and target products are characterized at the same time, so as to obtain the source domain user preference representation that takes into account the preference characteristics of the target product; based on the preference representation, a personalized mapping module is used to map the user representation from the source domain to the target domain, so as to obtain the user representation corresponding to the target domain.

3. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 2 is characterized in that: The dual-channel attention module includes Attn based on the attention mechanism src (·) and Attn tgt-item (·) Two feature encoders, for the current user Its preference feature representation is obtained by weighted fusion of its interacted product representations: Attn src (·) represents the corresponding attention network module, θ src represents the corresponding model parameters, Indicates user The lth interacted item in the source domain.

4. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 3 is characterized in that: The current user The embedded representation of the products that have interacted in the source domain is concatenated with the embedded representation of the target product. The concatenated representation sequence is used as the input of the attention module and the corresponding attention weights are calculated for weighted fusion to obtain the current user's preference representation, which is specifically: Attn tgt-item (·) represents the corresponding attention network module, θ tgt-item represents the corresponding model parameters, Indicates user The lth interacted item in the source domain, Indicates the target product.

5. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 4, characterized in that: The personalized mapping module is used to model the personalized preference characteristics of users and commodities, specifically: and The source domain user representation integrated with the target product information is mapped to the target domain using the personalized mapping module. The specific process is as follows: where g(·) represents the meta-network, represents the learnable parameters corresponding to the meta-network; is the output vector of the meta-network, whose size depends on the corresponding mapping function structure. The shape of is transformed into d×d to obtain the parameters of the personalized mapping module, Will As a parameter of the personalized mapping module, we get Representation in the target domain 6. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 5, characterized in that: In the public bridge module, all users use a common mapping function, focusing on the common features of the mapping from all users in the source domain to the target user, which is specifically: the public mapping function is directly represented by the source domain user As input, all users share common bridge parameters, and output It is a table of the user in the target domain after the source domain user is mapped through the public bridge and integrated with the user's public preference features. The specific formalization is as follows: Among them, f c (·) Using a two-layer perceptron structure, is the embedding representation of user i in the source domain, and ω is a learnable parameter.

7. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 6, characterized in that: The dynamic target perception fusion module can model the characteristics that change with the change of users, especially target products. The specific process is as follows: S1: Get the corresponding Q, K and V in the attention mechanism, Representation in the source domain Respectively and Splice and get and S2: Re-stacking and get Will As K and V in the attention mechanism; S3: Target product representation v t Respectively and After splicing and stacking, we get Use it as an attention mechanism.

8. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 7, characterized in that: The specific process of the dynamic target perception fusion module is as follows: Perform attention fusion output to obtain Among them, sm(·) represents the softmax function, and is a learnable parameter; Will Split into two d-dimensional vectors by row, using and Indicates that for The corresponding output after the dynamic target perception fusion module; for The corresponding output after the dynamic target perception fusion module.

9. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 8, characterized in that: The dynamic target-aware fusion module is used to explicitly model the features that change with the user and the target product, and obtain a representation that further fuses the features of the source domain user and the target product. and Will and Perform mean fusion to obtain the final user representation pass and The inner product of Target products Rating: The final objective function is: in, The set of true ratings of public users in the target domain, Score the model predictions.

10. The dynamic cross-domain recommendation method of target-aware complementary knowledge transfer and information fusion as claimed in claim 1, characterized in that: The model DnmCDR is a mapping-based cross-domain recommendation model.

Images