A cross-domain recommendation method and system based on a personalized preference transfer model

Abstract

The application discloses a cross-domain recommendation method and system based on a personalized preference transfer model, and the method comprises the following steps: constructing a personalized preference transfer model, first constructing a multi-relation graph according to the interaction, basic information and corresponding comments of a user and an item, and calculating the embedding of the user and the item in a source domain and a target domain according to the information in the graph; then identifying similar users of a target user in the source domain according to the learned user representation, learning a personalized preference transfer function by using the individual characteristics of the target user and the common characteristics of the similar users of the target user; finally transferring the user preference according to the learned personalized preference transfer function, calculating the predicted score of the target user to the items in the target domain according to the transferred user embedding, and realizing the personalized recommendation of the user in the source domain in the target domain; obtaining the interaction, basic information and corresponding comments of the user and the item and inputting the personalized preference transfer model to realize the personalized recommendation. The application provides better recommendation performance for cross-domain recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation systems, and in particular to a cross-domain recommendation method and system based on a personalized preference transfer model. BACKGROUND

[0002] Recommendation systems have been widely applied in the Internet field, which recommend personalized items to users, thus providing great help to solve the problem of information overload faced by users. Traditional recommendation systems use collaborative filtering (CF) and matrix factorization (MF) techniques to realize recommendation based on historical user-item interactions (such as rating, purchase or click behavior, etc.). The accuracy of the recommendation is largely determined by the number of user-item interactions. However, when dealing with new users, the interaction data becomes very sparse, leading to the so-called cold start problem of users.

[0003] In order to alleviate the cold start problem of users, many recent works have been proposed and achieved good results. These works either learn the preferences of cold start users from auxiliary information (such as user profiles and item attributes), or apply meta-learning to adapt the globally shared prior (or preference) to the individualized prior of each cold start user according to the sparse interaction data of each cold start user. Cross-domain recommendation (CDR) is another important technical route to alleviate the cold start problem of users, which aims to transfer useful knowledge from the source domain (or called auxiliary domain) to the target domain, and has recently attracted increasing attention from both academia and industry. Compared with those solutions based on auxiliary information or meta-learning, CDR is more challenging because it needs to reasonably solve two core problems, i.e., "what to transfer" and "how to transfer".

[0004] In order to solve the first core problem, classical machine learning methods such as factorization machines (FM) and latent semantic analysis (LSA) are first used to learn the useful knowledge (such as user preferences) of the auxiliary domain, and then transferred to the target domain to enhance the prediction of the user preferences of the target domain. Benefiting from the rapid development of deep learning technology, many deep learning-based methods have been proposed in recent years to solve the first core problem, and these methods are more powerful than classical machine learning methods, promoting the further development of CDR. In order to solve the second core problem, many works use the information of overlapping users (or items) to establish the connection between the two domains, but when there are no overlapping objects between the two domains, the connection between the two domains is established by capturing label correlation and active learning.

[0005] Although the existing research works have made significant progress, there are still the following shortcomings in them:

[0006] One is the neglect of user-user and item-item relationships when computing the representations of users and items. For the first core problem of CDR, i.e., "what to transfer", the most concerned task is to learn the representations of users and items in each domain, because the quality of these information plays a crucial role in the subsequent stages. Existing works only compute the representations of users and items based on the explicit user-item interactions, while the implicit user-user relationships (determined by user profiles and review information) and item-item relationships (by analyzing the descriptions of items) are usually neglected. Recent studies show that user-user and item-item relationships are very helpful to capture the characteristics of users and items comprehensively and accurately, especially for those users and items with fewer interactions. For example, Figure 1 Three heterogeneous relationships, i.e., user-item, user-user, and item-item relationships in the movie domain, are illustrated. As shown in Figure 1 , a cold-start user u i has no interactions with movies, which leads to a poor personalized recommendation effect for the user if only considering the user-item relationship. However, when considering the user-user and item-item relationships, the rating of the movie m i by the user u j can be better predicted. Specifically, since the users u2 and u4 with similar preferences to u i give the highest ratings of 5 to the movies m1 (Titanic) and m2 (Before Sunrise) respectively, while m j (Your Name) is very similar to m1 and m2 in terms of their tags (e.g., warm and romantic). In summary, the neglect of user-user and item-item relationships makes most of the recent studies less accurate in modeling the representations of users and items, which negatively affects the effectiveness of personalized recommendation in the target domain.

[0007] Two is that the individual characteristics and shared characteristics of users are not considered simultaneously when transferring user preferences. For the second core problem of CDR, i.e., "how to transfer", one of its main tasks is to transfer user preferences between two connected domains. To transfer user preferences, some research works consider that all users have the same relationship between the source domain and the target domain, and learn a general user preference transfer function for all users, as shown in Figure 2 (a). However, due to the complex personal characteristics of users, the general transfer function can reduce the performance of CDR. To overcome this shortcoming, Zhu et al. proposed to adopt a personalized preference transfer function (i.e., f1,..., f n ) based on the individual interactions of each user in the source domain, as shown in Figure 2(b) shown. However, the effectiveness of this approach heavily relies on the number of user-item interactions, which is very scarce in the cold-start scenario of users. In this case, according to the theory of user collaborative filtering, the common features extracted from similar users can be applied to improve the learned personalized transfer function. In this paper, we propose a novel approach to learn a personalized transfer function for each user by simultaneously considering the individual features of the user and the common features of similar users. Figure 2 (c) for example, the figure shows the user u j The historical interactions with items are very few, which indicates that the individual features encoded by the interactions of the user u j are insufficient. In this case, in addition to the individual features of u j , the common features from a group of users with similar interests to the user u j (such as G1={u1, u2,...}) can help improve the learned personalized transfer function Since the existing CDR methods cannot simultaneously use the individual features and common features (hereinafter collectively referred to as transferable features) of the user to learn the personalized preference transfer function of the user, the recommendation results produced by these methods are not satisfactory.

[0008] Therefore, there is an urgent need for an effective method to solve the above problems. SUMMARY

[0009] The technical problem to be solved by the present application is that in view of the technical problems existing in the prior art, the present application provides a cross-domain recommendation method and system based on a personalized preference transfer model. In order to solve the first problem well, first, based on the heterogeneous explicit and implicit relationships between users and items, the high-quality representation of the user and the item to be transferred is calculated. Then, the individual features of the user and the common features of similar users are considered simultaneously, and a more effective personalized preference transfer function between the two domains is learned for the user, effectively alleviating the second problem. Finally, based on the learned personalized preference transfer function of each user, the user embedding transferred to the target domain is calculated, realizing the personalized recommendation of each user in the target domain, and providing better recommendation performance for cross-domain recommendation.

[0010] To solve the above technical problems, the technical scheme provided by the present application is:

[0011] A cross-domain recommendation method based on a personalized preference transfer model, comprising the following steps:

[0012] S1) Construct a personalized preference transfer model. First, the model builds a multi-relationship graph based on user and item interactions, basic information, and corresponding comments. Then, based on the information in the multi-relationship graph, it calculates the embeddings of all users and items in both the source and target domains. Next, based on the learned user representations, it uses soft clustering in the source domain to identify similar users for each target user. It then uses the individual characteristics of the target user and the common characteristics of its similar users to learn the personalized preference transfer function for the target user. Finally, it calculates the user embedding transferred from the source domain to the target domain based on the learned personalized preference transfer function for each user. Finally, it calculates the predicted rating of the user for items in the target domain based on the transferred user embedding to achieve personalized recommendations for users in the source domain within the target domain.

[0013] S2) Obtain user and project interactions, basic information, and corresponding comments, and input them into the personalized preference transfer model to obtain personalized recommendations for each user.

[0014] Furthermore, in step S1, when constructing a multi-relationship graph based on user and item interactions, basic information, and corresponding comments, and calculating the embeddings of all target users and items in the source and target domains respectively based on the information in the multi-relationship graph, the specific steps include:

[0015] S101) Convert user-related basic information and user comments into user document vectors, and convert project-related basic information and user comments into project document vectors. Calculate the similarity probability between each pair of users in the user document vectors and the similarity probability between each pair of projects in the project document vectors. Then, aggregate all users and projects into nodes of a multi-relationship graph. Generate corresponding user-user heterogeneous edges in the multi-relationship graph based on the similarity probability between each pair of users, generate corresponding project-project heterogeneous edges in the multi-relationship graph based on the similarity probability between each pair of projects, and generate corresponding user-project edges in the multi-relationship graph based on the historical interactions between users and projects.

[0016] S102) Based on the embedding strategy oriented towards heterogeneous relationships, user-user relationships or item-item relationships in the multi-relationship graph are treated as similarity relationships, and user-item relationships in the multi-relationship graph are treated as interaction relationships. For each similarity relationship, Euclidean distance is used to measure the distance d between two nodes in the latent vector space. sr (n i ,n j ), and calculate d sr (n i ,n j The loss function at its minimum It also uses an explicit translation-based approach to model all interactions as translations d between nodes in the latent vector space. ir (n p ,n q), and calculate d ir (n p ,n q )min Finally, the loss function and the loss function are jointly minimized to obtain the embedding matrix of users or items.

[0017] Further, in the step S101, the corresponding user-user heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of users, the corresponding item-item heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of items, and the corresponding user-item edges in the multi-relation graph are generated according to the historical interaction between users and items, and the step S101 comprises:

[0018] If the similarity probability between a pair of users or the similarity probability between a pair of items is greater than a preset threshold, the corresponding user-user heterogeneous edge or item-item heterogeneous edge in the multi-relation graph is generated, and the similarity probability corresponding to the user-user heterogeneous edge or item-item heterogeneous edge is taken as the weight of the user-user heterogeneous edge or item-item heterogeneous edge.

[0019] If there is a historical interaction between users and items, the corresponding user-item edge in the multi-relation graph is generated, and the quotient of the historical rating of the user to the item and the maximum value in the preset rating matrix is calculated as the weight of the user-item edge.

[0020] Further, in the step S1, according to the learned user representation, similar users are identified for each target user in the source domain using a soft clustering method, and the personalized preference transfer function of the target user is learned using the individual features of the target user and the common features of the similar users of the target user, and the step S1 specifically comprises:

[0021] S201) Group all users in the source domain into K clusters, respectively calculate the allocation probability of the embedding of each user to the kth cluster, and then establish a target distribution Y to guide the optimization of the unsupervised clustering loss function , and further update the cluster center, and finally multiply the allocation probability of the target user to each cluster by the cluster center of the class, and then accumulate to obtain the common features of the target user.

[0022] S202) Obtain the interaction item list of the target user in the source domain, use an attention network to calculate the attention score of each item in the interaction item list, and normalize to obtain the weight of each item, multiply the embedding of each item in the interaction item list of the target user by the corresponding weight, and then accumulate to obtain the individual features of the target user.

[0023] S203) connecting the common features and individual features of the target user through a connection operation to obtain transferable features of the target user, using a preset neural network to learn a personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user, and then generating a representation of the target user after transfer in the target domain according to the personalized preference transfer function of the target user.

[0024] Further, in step S201, the target distribution Y is established to guide the optimization of the unsupervised clustering loss function , and further update the clustering centers, including:

[0025] Construct a soft assignment matrix that records the soft assignment information of K clusters of the source domain The definition of the target distribution Y is as follows:

[0026]

[0027] Where, ∑ i S i,k is the soft clustering frequency of the kth clustering center o k , S i,k represents the probability of assigning user u i to the kth cluster o k ;

[0028] The clustering loss function is defined by the KL divergence between the soft assignment matrix S and the target distribution Y Make the soft assignment close to the target distribution as follows:

[0029]

[0030] In each iteration, the set of clustering centers O = [o1, o2,..., o K ] is updated by stochastic gradient descent, and each clustering center o k corresponding Gradient calculation is as follows:

[0031]

[0032] Where, β represents the degrees of freedom of the t-distribution of the student, e ui represents the embedding of user u i , and the update of the kth clustering center is as follows:

[0033]

[0034] Where, ω is the learning rate that controls the update speed of the clustering center.

[0035] Further, the attention score expression of each item in the interaction item list in step S202 is as follows:

[0036]

[0037] where g(·; η) is an attention network, η is a learnable parameter, e vj denotes the embedding of item v j .

[0038] The weight expression of each item is as follows:

[0039]

[0040] where B ui denotes the list of items that user u i interacted with in the source domain, v l denotes the l-th item in the list of items that user u i interacted with in the source domain.

[0041] Further, in step S203, the preset neural network is used to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user, and then the representation of the target user after transfer in the target domain is generated according to the personalized preference transfer function of the target user, comprising:

[0042] The expression of using the preset neural network to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user is as follows:

[0043]

[0044] where vector contains the parameters of the personalized preference transfer function, and h(·; φ) is a two-layer neural network parameterized by φ, denotes the transferable features of the target user.

[0045] The vector is reshaped into a matrix to fit the size of the parameters of the preference transfer function, where d e denotes the embedding dimension of the user.

[0046] The matrix M ui is used as a parameter to generate the representation of the target user u i after transfer in the target domain by using the personalized preference transfer function of the target user u i , and the expression is as follows:

[0047]

[0048] where denotes the representation of the target user u i in the source domain.

[0049] Further, the step S1 includes the following steps: performing dot product operation on the representation of the target user after the transfer and the representation of each item in the target domain, respectively, to obtain the predicted score of the target user for each item in the target domain, and generating personalized recommendation for the target user according to the predicted score of each item.

[0050] Further, after obtaining the predicted score of the target user for each item in the target domain, the method further includes the following steps: training the personalized preference transfer function by using a task-based optimization method, and minimizing the loss between each predicted score and a real score.

[0051] The application further provides a cross-domain recommendation system based on the personalized preference transfer model, comprising a computer device programmed or configured to execute any of the cross-domain recommendation methods of the personalized preference transfer model.

[0052] Compared with the prior art, the application has the following advantages:

[0053] The application firstly mines the implicit user-user and item-item relationships according to the basic information and corresponding comments, constructs a heterogeneous information network containing user and item nodes and three types of edges (user-item, user-user and item-item) in each domain as a multi-relation graph, and calculates more representative user and item embeddings of the source domain and the target domain based on the information in the multi-relation graph, thereby generating rich representations of users and items based on the graph, and effectively solving the limitation of the prior art that the user-user relationship and the item-item relationship are ignored when calculating the representations of users and items.

[0054] The application considers the individual characteristics of the user and the common characteristics of similar users, uses a soft clustering method to identify similar users for each target user, and uses the individual characteristics of the user and the common characteristics of similar users to learn the personalized preference transfer function of the user, thereby learning a more effective personalized preference transfer function between the two domains, and effectively solving the limitation of the prior art that the individual characteristics and the common characteristics of the user are not considered simultaneously when transferring the user preference. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A schematic diagram for inferring the preference of a user for an item by considering three heterogeneous relationships.

[0056] Figure 2 A schematic diagram of the three different user preference transfer methods.

[0057] Figure 3 A flowchart of the embodiment of the application.

[0058] Figure 4 A schematic diagram of the framework of the personalized preference transfer model in the embodiment of the present application.

[0059] Figure 5 A control diagram of the influence of parameter setting on performance of the personalized preference transfer model in the embodiment of the present application.

[0060] Figure 6 A control diagram of the visualized test results of the personalized preference transfer model and the existing model in the embodiment of the present application. DETAILED DESCRIPTION

[0061] The present application will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present application is not limited thereby.

[0062] Before introducing the specific embodiments of the present application, the related concepts are introduced first:

[0063] User cold-start problem: Traditional recommender systems use CF techniques to learn user preferences based on their historical interactions, thus cannot fully handle cold-start users with sparse interactions in the system. To alleviate the user cold-start problem, most existing solutions suggest directly leveraging auxiliary information, such as enriching user and item representations by user profiles and item attributes. Recent studies use HINs to model heterogeneous relationships between different objects and use high-order auxiliary information of users or items captured from HINs to further address the user cold-start problem. Meanwhile, the concept of "learning to learn", i.e., the meta-learning paradigm, is also used to alleviate the user cold-start problem. This paradigm can quickly adjust the meta-learning prior (or preference) shared globally among all users based on the user's scarce interaction data to adapt to the personalized prior of each cold-start user. For example, MAMO and MetaHIN combine user profiles and constructed HINs with the meta-learning paradigm to alleviate the user cold-start problem at the data level and model level, respectively. In addition, PAML and CBML suggest using local priors learned from similar users in the meta-learning process instead of global priors to avoid the negative impact of unrelated users. Cross-domain recommendation (CDR) is another important technique to address the user cold-start problem. In CDR solutions, the knowledge of the information auxiliary domain is transferred to the target domain to enrich the representation of users or items and improve the recommendation performance of cold-start users. Available surveys show that CDR is more complex than the auxiliary information-based or meta-learning-based strategies mentioned earlier, as it needs to consider various recommendation scenarios (user overlapping scenarios or non-overlapping scenarios) and different recommendation tasks (such as intra-domain recommendation or inter-domain recommendation). In addition, CDR not only needs to model user preferences within a single domain, but also needs to consider transferring user preferences from the source domain to the target domain. Recently, CDR has attracted increasing attention from academia and industry to address the cold-start problem in recommendations.

[0064] Cross-domain recommendation: Cross-domain recommendation (CDR) aims to transfer useful knowledge from the source domain (or called auxiliary domain) to the target domain, which has two core problems, i.e., "what to transfer" and "how to transfer".

[0065] To solve the first core problem, EMCDR based on the historical user-item interaction data, adopts the traditional matrix factorization (MF) and Bayesian personalized ranking (BPR) methods to compute the representation of users and items based on the historical user-item interaction, which will also be transferred to the target domain in the subsequent stage. However, EMCDR does not use any auxiliary information, resulting in poor embedding effect of users and items. RC-DFM uses a variant of stacked denoising autoencoder (SDAE) to fuse auxiliary information (i.e., review texts and item content) with the user-item rating matrix and generate more rich representation for users and items. Meanwhile, many studies also utilize inferred auxiliary information, such as user-user or item-item relationships, to further improve the representation of users or items. For example, CDLFM determines user-user relationships through three pre-defined user similarity measures on user-item interactions and incorporates these relationships as constraints in the matrix factorization process. In DCDIR, the relationships among items are mined through the meta-paths defined on the insurance product knowledge graph, improving the effectiveness of the representation of items in the target domain. Recently, Zhu et al. established a heterogeneous graph that captures heterogeneous relationships (i.e., user-item, user-user, and item-item) in their recommendation (i.e., GA-DTCDR) to enhance the modeling of users and items. However, the use of homogeneous graph embedding techniques (i.e., Node2vec) in GA-DTCDR to compute the embedding of users and items cannot distinguish different relationships in the graph, reducing the quality of the computed embedding vectors. Therefore, it is necessary to design a heterogeneous graph embedding method to better compute the knowledge that needs to be transferred in the source domain based on heterogeneous relationships.

[0066] To address the second core issue, information about overlapping users or items in two domains is often used to establish connections between the two domains, while tag relevance capture and proactive learning are often used to establish connections between two domains when there are no overlapping objects. This study focuses on scenarios where users in two domains partially overlap, which is very common. In this case, the proposed CDR method assumes that the relationship between user preferences in the source and target domains is shared by all users and learns a general preference transfer function for all users. However, due to the complex individual characteristics of users, a general preference transfer function may degrade the performance of CDR. To overcome this drawback, CGN uses a recurrent generative network to develop a personalized bidirectional transfer function, while PTUPCDR uses a meta-network to generate a personalized transfer function for each user. However, these works only use each user's historical behavior (referred to as individual features in this paper) to train their models, which leads to unsatisfactory recommendation results when dealing with cold-start users. In the cold-start user scenario, shared features from a group of users with similar preferences can significantly improve the quality of the personalized transfer function learned by the user. However, existing CDR methods do not simultaneously apply both individual and shared features of users to learn the personalized preference transfer function, resulting in less accurate recommendations.

[0067] Example 1

[0068] like Figure 3 As shown, this embodiment proposes a cross-domain recommendation method based on a personalized preference transfer model, including the following steps:

[0069] S1) Construct a personalized preference transfer model, the personalized preference transfer model as follows: Figure 4 As shown, firstly, a multi-relationship graph is constructed based on user and item interactions, basic information, and corresponding comments. Then, embeddings of all users and items are calculated in both the source and target domains based on the information from the multi-relationship graph. Next, based on the learned user representations, soft clustering is used in the source domain to identify similar users for each target user. The individual characteristics of the target user and the common characteristics of its similar users are used to learn the personalized preference transfer function for the target user. Finally, the user embedding transferred from the source domain to the target domain is calculated based on the learned personalized preference transfer function for each user. Based on the transferred user embeddings, the predicted ratings for items in the target domain are calculated to achieve personalized recommendations for users in the source domain within the target domain.

[0070] S2) Obtain user and project interactions, basic information, and corresponding comments, and input them into the personalized preference transfer model to obtain personalized recommendations for each user.

[0071] Before describing the specific steps, we first present the formulation of our problem and then define the key concepts related to our problem.

[0072] In this embodiment, we denote the set of users as U = {u1, u2, u3, … u N and the set of items as V = {v1, v2, v3, … v M . In addition, the rating matrix between users and items can be denoted as R, where r i,j ∈R represents the rating between user u i and item v j . Note that there are two domains in each CDR task, and we use superscript {s, t} to distinguish them. For the partial user overlap CDR scenario, there are partial users who have interactions in both domains, while the items in the two domains are disjoint. Therefore, we call the overlapping users as U o = U s ∩U t , where |U o | << |U s | or |U t |. The purpose of this embodiment is to optimize the rating of user u i to item v j in the target domain

[0073] To obtain high-quality user and item embeddings, explicit user-item and implicit user-user and item-item relationships are obtained by inference on the multi-relational graph constructed in the proposed personalized preference transfer model (HCCDR). It should be noted that the multi-relational graph is a heterogeneous information network, which contains three heterogeneous relationships (user-item, user-user, item-item), and each relationship in the graph is represented as a node-relation triple in this embodiment. Then, a method for computing user and item embedding vectors based on information in the multi-relational graph is proposed. We define the relevant key concepts as follows.

[0074] Definition 1 (Multi-relational graph): A multi-relational graph is represented as where each node n ∈ V and each edge e ∈ E. Each multi-relational graph is associated with a node type mapping function and an edge type mapping function . and T E describe the set of node and edge types, respectively, satisfying

[0075] Definition 2 (Node-relation triple): In a multi-relational graph, each relationship is represented as a node-relation triple, which is in the form of <n i ,e,n j> indicates that a relation (i.e., an edge) e∈E connects two nodes n. i ,n j ∈V, and each node-relation triplet <n i ,e,n j >∈P, where P is the set of all node-relation triples.

[0076] Definition 3 (Heterogeneous Information Network Embedding): Given a multi-relationship graph There exists a mapping function Where d << |V| is trained to project each node n ∈ V into a low-dimensional vector space.

[0077] We have summarized the key symbols and their descriptions in Table 1.

[0078] Table 1 Key Symbols and Explanations

[0079]

[0080]

[0081] In this embodiment, the framework of the personalized preference transfer model constructed in step S1 is as follows: Figure 4 As shown, it consists of three parts: a heterogeneous latent factor modeling (HLFM) component, a cluster-enhanced preference transfer (CPT) component, and a personalized recommendation (PR) component. Figure 4 in, u i It is one of the overlapping users between the source and target domains, and v j This refers to an item within the target domain. Specifically, the heterogeneous latent factor modeling component is responsible for calculating the embeddings of users and items in each domain. Then, the clustering-enhanced preference transfer component considers both common and individual user characteristics to learn a personalized preference transfer function for each user. Finally, the personalized recommendation component implements personalized recommendations in the target domain based on the user's preferences as they transition from the source domain to the target domain. Therefore, step S1 in this embodiment, constructing the personalized preference transfer model, includes: constructing a heterogeneous latent factor modeling component, constructing a clustering-enhanced preference transfer component, and constructing a personalized recommendation component.

[0082] For heterogeneous latent factor modeling components, as mentioned earlier, in addition to explicit user-item relationships generated based on historical user-item interactions, implicit user-user relationships and item-item relationships also play a crucial role in improving user and item representations. By considering these heterogeneous relationships, we designed a Heterogeneous Latent Factor Modeling (HLFM) component to learn user and item representations. In HLFM, user-item relationships are first identified. Then, a multi-relationship graph is constructed for each domain, containing two types of nodes (user and item nodes) and three types of relationship edges (user-item, user-user, and item-item). Finally, we designed an embedding method oriented towards heterogeneous relationships to compute enhanced representations of users (or items) in the form of embedding matrices. To achieve the above functionality, HLFM mainly consists of two modules: a multi-relationship graph construction module and a heterogeneous relationship-oriented embedding module.

[0083] For multi-relationship graph construction modules, users (or projects) are typically interconnected. Relationships between users (or projects) can be identified based on their basic profiles, such as shared educational backgrounds (or manufacturers) and related user reviews, reflecting user interests or project characteristics. In this paper, c i (c N+j ) indicates storage and user u i Documentation containing relevant basic information and user reviews, where 1 ≤ i ≤ N (or item v) j Where 1≤j≤M). In this embodiment, all variables formed by C={c1,c2,…,c N+M The documents represented by} were all segmented into words using the popular natural language tool Stanford CoreNLP. Then, the document vectors of all users (i.e., ) and the document vectors (i.e., D) of all items V =[d vj ] M×dc Both ) can be calculated using the widely used Doc2vec, where d c It is the dimension of a document vector. The document vector derived from all users, i.e., D... U Similarity probability between a pair of users i u j The calculation is as follows:

[0084]

[0085] Among them, S(du) i , du j ) for u i Document vectors (i.e., du) i ∈D U ) and u j Document vectors (i.e., du)​j ∈D U ) between 0 and 1. a is a hyper-parameter that controls the similarity threshold. If the value of S(·, ·) exceeds the threshold a, the output of T is equal to the input, otherwise the output of T is set to 0. Similarly, the similarity probability between a pair of items can also be computed.

[0086] Then, all users and items are summarized as nodes and three types of heterogeneous edges (i.e., user-item, user-user, and item-item) are generated to construct a multi-relational graph. Specifically, a user-item edge is generated if there exists a historical user-item interaction, and the weight of the edge is computed by r i,j / max(R). Here, r i,j is the rating of user u i to item v j , and max(R) is the maximum value in the rating matrix R. If the similarity probability of a user-user (or item-item) edge is greater than 0, it is generated, and the weight of the edge is set to the similarity probability.

[0087] For the heterogeneous relation-oriented embedding module, after constructing the multi-relational graph, we implement a heterogeneous relation-oriented embedding (HRE) strategy to compute the embedding matrix of users (or items) based on the graph. The multi-relational graph G contains three types of heterogeneous relations (i.e., user-item, user-user, and item-item); the embedding strategy should consider their different semantics. To facilitate subsequent modeling, we first treat a user-user (or item-item) relation as a similarity relation (SR), as it reflects the similarity between two users (or items) in their attributes. Meanwhile, we treat a user-item relation as an interaction relation (IR), which represents the user-item interaction. In this embodiment, the HRE strategy is used to handle the similarity relations (SRs) and interaction relations (IRs) in the multi-relational graph, using different ways to better capture their semantic differences.

[0088] For SR, HRE ensures that two similar users (or items) in the participating relation are mapped close to each other in the latent vector space. The Euclidean distance is used in HRE to measure the distance between two nodes in the latent vector space. Given a node-relation triple <n i , e, n j >, the distance between nodes n i and n j in the latent vector space is calculated as follows:

[0089]

[0090] where w i,j is the weight of edge e (or relation), and Hn i are the d-dimensional embedding vectors of node n i and node n j respectively. To ensure that two nodes are close to each other in the latent vector space, we adopt an edge-based loss function to minimize d sr (n i ,n j ), as follows:

[0091]

[0092] where γ is a marginal hyper-parameter, we set γ = 1. P SR is the set of node-relation triplets belonging to SRs, while is the negative sample set of node-relation triplets belonging to SRs.

[0093] Since IRs convey the interaction information between user nodes and item nodes, HRE uses an explicit translation-based method to model IRs as translations between nodes in the latent vector space. Formally, given a node-relation triplet <n p , r, n q >, r belongs to IR, the distance between nodes n p and n q is defined as follows:

[0094]

[0095] where w p,q is the weight of relation r, Hn p , Hn q are the embeddings of nodes n p and n q respectively, and X r is the embedding of relation r. Equation (4) penalizes the deviation of (Hn p + X r ) from Hn q . Similarly, an edge-based loss function is defined to ensure that the translation between two nodes in the low-dimensional vector space:

[0096]

[0097] where P IR is the set of node-relation triplets belonging to IRs, while is the negative sample set of node-relation triplets not belonging to IRs.

[0098] Finally, the two loss functions above are jointly minimized as follows, generating the embedding matrix of users (or items), i.e., E U = [e ui ]​N×de (or E V = [e vj ] M×de ), where d e is the dimension of the embedding vector of the user (or item).

[0099]

[0100] Therefore, based on the heterogeneous potential factor modeling component constructed in the embodiment, in step S1 of the embodiment, the personalized preference transfer model constructs a multi-relation graph (user-item relation, user-user relation, item-item relation) according to the interaction, basic information and corresponding comments of the user and the item, and calculates the embedding of all users and items in the source domain and the target domain according to the information of the multi-relation graph, including the following steps:

[0101] S101) Based on the multi-relation graph construction module, the user-related basic information and user comments are converted into user document vectors, and the item-related basic information and user comments are converted into item document vectors. The similarity probability between each pair of users in the user document vector is calculated, and the similarity probability between each pair of items in the item document vector is calculated. Then all users and items are summarized as nodes of the multi-relation graph, the corresponding user-user heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of users, the corresponding item-item heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of items, and the corresponding user-item edges in the multi-relation graph are generated according to the historical interaction between the user and the item;

[0102] S102) Based on the embedding module for heterogeneous relations, according to the embedding strategy for heterogeneous relations, the user-user relation or the item-item relation in the multi-relation graph is regarded as a similar relation, and the user-item relation in the multi-relation graph is regarded as an interaction relation. For each similar relation, the Euclidean distance d sr (n i ,n j ) between two nodes in the potential vector space is measured, and the loss function sr (n i ,n j ) is calculated, which is the minimum of d All interaction relations are also modeled as translation d i (n p ,n q ) between nodes in the potential vector space using an explicit translation-based method, and the loss function i (n p ,n q ) is calculated, which is the minimum of d Finally, the loss function and the loss function Co-minimization, resulting in the embedding matrix of users or items.

[0103] In step S101 of the embodiment, the corresponding user-user heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of users, the corresponding item-item heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of items, and the corresponding user-item edges in the multi-relation graph are generated according to the historical interaction of users and items.

[0104] If the similarity probability between a pair of users or the similarity probability between a pair of items is greater than zero, the corresponding user-user heterogeneous edge or item-item heterogeneous edge in the multi-relation graph is generated, and the similarity probability corresponding to the user-user heterogeneous edge or item-item heterogeneous edge is taken as the weight of the user-user heterogeneous edge or item-item heterogeneous edge.

[0105] If there is a historical interaction of users and items, the corresponding user-item edge in the multi-relation graph is generated, and the quotient of the historical rating of the user to the item and the maximum value in the preset rating matrix is calculated as the weight of the user-item edge.

[0106] For the cluster-enhanced preference transfer component, we believe that providing personalized transfer functions for each user is very important to ensure the performance of CDR. Personalized transfer functions can be learned from user preferences. Although the interaction between users and items directly reflects the individual characteristics of user preferences, such interaction captured from the source domain is not enough in the user cold start setting. In order to better learn the personalized transfer function, a cluster-enhanced preference transfer (CPT) component is proposed to transfer the target user's preferences from the source domain to the target domain based on the individual characteristics of the user and the common characteristics learned from similar users. Specifically, first, a soft clustering method is applied in CPT to identify similar users of the target user. Therefore, it is possible to derive the common characteristics of the user's preferences compared with similar users, which is considered as an important supplement to individual characteristics. Finally, through these two characteristics of user preferences, a high-quality personalized transfer function can be learned for each user. CPT has three main modules: common characteristic generator, individual characteristic generator, and user preference transfer.

[0107] For the common characteristic generator (CCG) module, a soft clustering method is used in the embodiment to identify a group of users with similar preferences, and then generate the common characteristics of the users according to the clustering results. Please note that soft clustering is also known as fuzzy clustering, which allows each user to be assigned to several classes with different assignment probabilities at the same time.

[0108] Specifically, the CCG module groups all users in the source domain into K clusters, and the kth class is identified by a cluster center, i.e. k Here, let the randomly initialized denotes the set of K cluster centers. Then, the student's t-distribution is adopted to facilitate the computation of the probability s i of assigning user u ui to the k-th cluster o U . k i,k i,k The computation of s i is referred to formula (7):

[0109]

[0110] where β is the degree of freedom of the student's t-distribution, usually set to 1.

[0111] Similar to the DEC method, unsupervised learning is used in the CCG module to optimize the clustering loss, denoted as Specifically, we construct a soft assignment matrix S which records the soft assignment information of the K clusters of the source domain. Then, the target distribution Y is established to guide the optimization of the unsupervised clustering loss and further update the cluster centers. The definition of the target distribution Y is as follows:

[0112]

[0113] where ∑ i S i,k is the soft clustering frequency of the k-th cluster o k S i,k denotes the probability of assigning user u i to the k-th cluster o k .

[0114] After the calculation of Y, the KL divergence between the soft assignment S and the target distribution Y is used to define the clustering loss function to make the soft assignment close to the target distribution as follows:

[0115]

[0116] In each iteration, the set of cluster centers O = [o1, o2,..., o K ] is updated by stochastic gradient descent (SGD). Each cluster center o k corresponds to gradient calculation as follows:

[0117]

[0118] The update of the k-th cluster center is shown as follows:

[0119]

[0120] where ω is the learning rate that controls the speed of updating the cluster centers.

[0121] Finally, the common feature of the target user u i is obtained by multiplying the assignment probability of the user to each cluster by the cluster center of the cluster and then accumulating the results.

[0122]

[0123] where s i,k is the probability of assigning the user u i to the k-th cluster o k calculated by formula (7).

[0124] In summary, the CCG module divides all users into K groups according to the user embeddings, and generates the common feature of each user according to the clustering results. Compared with the traditional hard clustering method, the soft clustering used in the CCG module is more effective, because it learns the assignment with high confidence with the help of the target distribution Y, iteratively improving the initial assignment results.

[0125] For the individual feature generator (ICG) module, in order to generate the individual feature of the user, the individual feature generator (ICG) module first adopts the attention mechanism to weight all the items of user interaction, reflecting the user's preference in the source domain. In form, set B ui = {v1, v2,...} is the list of interaction items of user u i in the source domain. The attention score of each item is learned by the attention network defined as follows:

[0126]

[0127] where the function g(·; η) is the attention network, and η is the learnable parameter. Please note that in this paper, the attention network is a two-layer MLP. Then, the normalized attention score a j of the item v j , i.e. the weight (or contribution) of the item v j , is calculated as follows:

[0128]

[0129] Finally, the individual feature of the user u i is calculated by the ICG module by aggregating the embeddings of all interaction items of the user u i , i.e. , with the weight as follows:

[0130]

[0131] For the User Preference Transfer module (UPT), to compute the personalized user preference transfer function of user u i , we first concatenate the common features (i.e. ) and the individual features (i.e. ) of user u i , forming the transferable features of user u , denoted as , where is the concatenation operation. Then, we design a neural network with the transferable features of user u i in the UPT module, learning the personalized preference transfer function of user u i from the source domain to the target domain, as follows:

[0132]

[0133] where the function is a two-layer neural network parameterized by , and the vector contains the parameters of the preference transfer function.

[0134] Although the preference transfer function can be set to any structure, we set it to a linear layer f(·) in the UPT module for simplicity. Note that the resulting vector is reshaped to a matrix to fit the size of the parameters of the preference transfer function, where d e denotes the embedding dimension of users. With M ui as the parameter, we generate the representation of user u i after transferring in the target domain, using the derived personalized preference transfer function of user u i , as follows:

[0135]

[0136] where denotes the representation of user u i in the source domain, and is the embedding of user u i after transferring from the source domain to the target domain. The derived can be regarded as the initial embedding of user u i in the target domain, to solve the cold start problem of users.

[0137] As can be seen, the User Preference Transfer module (UPT) learns the personalized preference transfer function of each user according to the personal and common features of users, improving the quality of the learned transfer function compared to existing CDR solutions.

[0138] Therefore, based on the cluster-enhanced preference transfer component constructed in the embodiment, in step S1 of the embodiment, when the personalized preference transfer model learns the target user's personalized preference transfer function according to the learned user representation, identifies similar users for each target user in the source domain using a soft clustering method, and uses the individual features of the target user and the common features of the similar users of the target user to learn the personalized preference transfer function of the target user, the following steps are included:

[0139] S201) Based on the common feature generator module, group all users in the source domain into K clusters, respectively calculate the assignment probability of assigning the embedding of each user to the kth cluster, and then establish a target distribution Y to guide the optimization of the unsupervised clustering loss function , and further update the cluster center, and finally multiply the assignment probability of the target user to each cluster by the cluster center of the class, and then accumulate to obtain the common features of the target user;

[0140] S202) Based on the individual feature generator module, obtain the interaction item list of the target user in the source domain, use an attention network to calculate the attention score of each item in the interaction item list, and normalize to obtain the weight of each item., multiply the embedding of each item in the target user's interaction item list by the corresponding weight and accumulate to obtain the individual features of the target user;

[0141] S203) Based on the user preference transfer module, connect the common features and individual features of the target user through a connection operation to obtain the transferable features of the target user, use a pre-set neural network to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user, and then generate the representation of the target user after transfer in the target domain according to the personalized preference transfer function of the target user.

[0142] For the personalized recommendation component, in the embodiment, for the target user u i , given its transferred user representation , the predicted score of the item v j in the target domain is calculated as follows:

[0143]

[0144] wherein, is the item representation in the target domain, and ⊙ represents the dot product operation.

[0145] ​Since the number of overlapping users between the two domains is usually very limited, the training of the mapping-based personalized preference transfer function can suffer from overfitting. Therefore, in the clustering-enhanced preference transfer (CPT) component in the present embodiment, a task-based optimization method is employed to train the personalized preference transfer function, minimizing the loss between each predicted rating and the true rating, which increases the training samples and avoids overfitting. The predicted loss is represented as:

[0146]

[0147] where, is the set of ratings of items in the target domain by overlapping users of the two domains.

[0148] Finally, the overall loss function of the proposed HCCDR model is defined as the sum of the clustering loss and the predicted loss as follows:

[0149]

[0150] Therefore, based on the personalized recommendation component constructed in the present embodiment, in step S1 of the present embodiment, the step of calculating the predicted rating of the target user for the items in the target domain by the personalized preference transfer model after the transfer of the target user to the target domain includes: performing dot product operation on the transferred representation of the target user in the target domain and the representation of each item in the target domain, respectively, to obtain the predicted rating of the target user for each item in the target domain, and generating personalized recommendation for the target user according to the predicted rating of each item.

[0151] Further, when constructing the personalized recommendation component in the present embodiment, after obtaining the predicted rating of the target user for each item in the target domain, the method further includes: training the personalized preference transfer function in a task-based optimization manner, minimizing the loss between each predicted rating and the true rating, which can effectively increase the training samples and avoid overfitting.

[0152] The following through specific experiments for the effect of the personalized preference transfer model in the present embodiment.

[0153] 1. Experimental setup

[0154] To evaluate the effectiveness of the proposed personalized preference transfer model (HCCDR), we use two publicly available datasets, Douban dataset and Amazon dataset, which are well-suited for CDR experiments. Both datasets have rich items from different domains and provide users' reviews and ratings. Each rating value ranges from 1 to 5, with higher ratings indicating greater user interest in the item. Detailed information about the two datasets is provided below, and their statistical data is summarized in Table 2. (s) and (t) in Table 2 represent the source domain and the target domain, respectively.

[0155] Douban dataset: Douban is one of the most authoritative rating platforms, considering its wide range of users and real user-generated reviews. It provides rich information about users and items, including user profiles, item descriptions, and user ratings and reviews of items. In our setting, we selected the movie domain and the book domain of Douban to define the CDR task1: Movie→Book. We filtered out all missing and duplicate data.

[0156] Amazon dataset: The Amazon dataset is one of the most commonly used datasets for evaluating CDR tasks. It contains nearly 20 domains, and in each domain, the dataset provides user ratings and reviews, product metadata, images, and links. For the Amazon dataset, we selected two pairs of domains to define two CDR tasks, namely Sport→Clothing (task2) and CD→Game (task3).

[0157] Table 2 Statistical data of two datasets

[0158]

[0159] We use two metrics, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), to evaluate the proposed personalized preference transfer model. These metrics are commonly used to evaluate the performance of CDR tasks, where:

[0160] Mean Absolute Error (MAE) is the average of the absolute errors between all predicted values and their corresponding ground truths. It reflects the actual error of the predicted values. The smaller the MAE value, the closer the predicted values are to their ground truths.

[0161] Root Mean Square Error (RMSE) is used to measure the deviation between all predicted values and their corresponding ground truths. It calculates the average of the squared errors between predicted values and ground truths, and then the RMSE is equal to the square root of the average. Due to the squaring operation, the RMSE value is usually greater than or equal to the MAE value and is sensitive to outliers. The smaller the RMSE value, the more accurate the predicted values.

[0162] We compare our proposed personalized preference transfer model with two types of baseline methods: 1) single-domain recommendation methods, including MF and GMF; 2) cross-domain recommendation methods, including CMF, EMCDR, DCDCSR, SSCDR, DCDIR, LACDR, and PTUPCDR. Detailed descriptions of the above baselines are as follows:

[0163] MF is a classical recommendation method that decomposes the sparse user-item interaction matrix into two low-rank and dense matrices, i.e., user latent factor matrix and item latent factor matrix, through which the rating score of users to items can be predicted.

[0164] GMF is a variant of neural collaborative filtering (NCF) model. When performing a dot product operation between user embeddings and item embeddings, it assigns different weights to each dimension to predict the rating.

[0165] CMF introduces different side information matrices to alleviate the cold start problem. It shares the user-item interaction matrix between the source domain and the target domain and jointly decomposes them. However, CMF ignores the differences between the two domains.

[0166] EMCDR is an embedding and mapping framework paradigm that trains a mapping function between the source domain and the target domain based on the information of overlapping users. By using the trained mapping function, the user embeddings of users who are more active in the source domain but have no behavior in the target domain are calculated, based on which personalized recommendations are made in the target domain.

[0167] DCDCSR generates a benchmark matrix using the matrix sparsity information between different domains, and then learns a mapping function between the target domain and the benchmark domain.

[0168] SSCDR believes that in real life, the proportion of common users or items between two domains is usually very small, resulting in poor mapping functions. Therefore, SSCDR learns the mapping function through non-overlapping data in a semi-supervised manner to enhance the robustness of the learned function.

[0169] DCDIR constructs a heterogeneous knowledge graph in the target domain and defines multiple meta-paths to obtain user embeddings before training the mapping function.

[0170] LACDR utilizes all user data to train the encoder of the source domain and the decoder of the target domain, respectively, achieving better generalization.

[0171] PTUPCDR believes that the preference transfer process of each user should be different. Therefore, it generates a personalized bridge for each user using individual interactions.

[0172] In the user preference transfer module of the proposed personalized preference transfer model, we employ a two-layer neural network with a hidden layer of 5×d. e Dimension, output layer is We employ the Adam optimizer with a learning rate of 0.005 and a batch size of 512. To ensure fairness in comparisons, for all baselines, we use the neural network-based model GMF to compute user and item embeddings while preserving the original optimal settings for other parameters. To evaluate the effectiveness of the proposed personalized preference transfer model for cold-start users, we split the data of overlapping users between the two domains into training and test sets, with ratios of 8:2, 5:5, and 2:8, respectively. We ignore user-item interactions in the test set and treat them as cold-start users in the target domain, while user ratings of items in the training set are used to train the personalized preference transfer function. We implement the personalized preference transfer model using PyTorch 1.10.0, which is trained using a GPU (Nvidia RTX A5000) and a CPU (Intel Core i9-11900 2.50G).

[0173] 2. Parameter Analysis

[0174] We first investigated the impact of different parameter settings on the performance of the personalized preference transfer model. The experimental results are as follows: Figure 5 As shown, Figure 5 In this model, the smaller the mean absolute error (MAE) or root mean square error (RMSE) value, the better the performance of the personalized preference transfer model (HCCDR).

[0175] Number of clusters K: from sub Figure 5 As shown in (a)-5(c), with the increase of the number of clusters K, the mean absolute error (MAE) or root mean square error (RMSE) of the personalized preference transfer model initially decreases steadily. This is because a small K leads to dissimilar users being grouped into the same cluster, thus reducing the quality of clustering. Then, when K is greater than 12 in Task 1, and greater than 10 in Tasks 2 and 3, the MAE or RMSE of the personalized preference transfer model increases. This is because a larger K results in a smaller number of users in the cluster, thus reducing the common feature information learned from users within the cluster. In the following experiments, we set K=12 for Task 1 and K=10 for Tasks 2 and 3 to ensure that the personalized preference transfer model achieves optimal performance.

[0176] Embedded dimension d e :son Figure 5(d)-5(f) describe the performance of the personalized preference transfer model as the embedding dimension d of users and items is simultaneously changed in the three tasks e These subplots show that as d e increases from 16 to 128, the performance of the personalized preference transfer model improves because d e larger generally means more user or item information can be encoded into their embedding vectors. At the same time, we can also observe from the subplots that when d e is larger than 32 (or when d e is larger than 64 in Task 2 and Task 3), the increase of d e brings smaller and smaller performance gain. And considering the fact that the computational overhead of the personalized preference transfer model grows exponentially with the increase of d e , in the following experiments, we always set d e = 32 for Task 1, and d e = 64 for Task 2 and Task 3.

[0177] Similarity threshold a: The value of similarity threshold a affects the quality of the computed embeddings of users or items in the process of constructing user-user relationships (or item-item relationships) in the personalized preference transfer model. When the value of a is changed, the performance of the personalized preference transfer model in the three tasks is shown in subplots Figure 5 (g)-5(i). These subplots reveal that as a increases, the mean absolute error (MAE) or root mean square error (RMSE) of the personalized preference transfer model first decreases and then increases because a larger a means only users (or items) with higher similarity are connected, thus the quality of the learned embedding vectors is improved. However, when a exceeds its optimal value, the number of generated relationships decreases, thus the information of the learnable user-user (or item-item) relationships is reduced, thus the quality of the learned embeddings of users or items is reduced. In the following experiments, a is set to 0.95 for Task 3, and a is set to 0.9 for Task 1 and Task 2.

[0178] 3. Comparison with baselines

[0179] Based on the three real datasets, we compare the personalized preference transfer model (HCCDR) with nine baseline methods on the three CDR tasks. The values in Table 3 are the average of the experimental results of five runs. The ratio represents the ratio of the training set to the test set. The best value is in bold, and the second best value is underlined. The improvement in the table is the improvement of HCCDR relative to the second value. From the results of Table 3, we can draw the following findings:

[0180] Discovery 1: Our proposed HCCDR outperforms all baselines in all three CDR tasks. In particular, the maximum performance improvement of HCCDR on the root mean square error (RMSE) (or mean absolute error (MAE)) is 12.69% (or 8.99%), and the minimum performance improvement on the RMSE (or MAE) is 3.38% (or 1.88%), which demonstrates the effectiveness of our method in handling CDR tasks.

[0181] Discovery 2: All single-domain recommendation methods (such as MF and GMF methods) are outperformed by these CDR methods (such as CMF, EMCDR, DCDCSR, SSCDR, DCDIR, LACDR, PTUPCDR, and HCCDR). This is because CDR methods leverage information from other domains, thereby alleviating the cold-start problem of users and improving recommendation performance.

[0182] Discovery 3: The improvement in target domain recommendation performance is related to the density of the source domain. For example, the average improvement in MAE and RMSE of HCCDR in Task 1 (source domain density = 4.047% as shown in Table 2) reaches 6.71%, while the average improvement in Task 2 (source domain density = 0.049%) and Task 3 (source domain density = 0.044%) is only 4.51% and 3.74%, respectively. The reason is that the higher the density of the source domain, the more user interaction information it contains, which helps to model user preferences in the target domain for better recommendations. Note that the density value of the source domain is the smallest in Task 3, which leads to the weakest improvement in the target domain.

[0183] Discovery 4: The larger the ratio of training set to test set, the better the recommendation performance. Specifically, when the ratio is set to 8:2, 5:5, and 2:8, respectively, the MAE or RMSE value of HCCDR gradually increases. This is because when the training set is larger, more training samples are used to train HCCDR, which makes HCCDR better generalize when handling recommendation tasks.

[0184] Discovery 5: CMF does not perform as well as other CDR methods (such as EMCDR, DCDCSR, SSCDR, DCDIR, LACDR, PTUPCDR, and HCCDR) in all three tasks. This is because CMF suggests combining data from the source and target domains to learn the preferences of overlapping users, which ignores the potential transfer between the two domains. In contrast, other CDR methods learn a preference transfer function to connect the two domains, which effectively alleviates the impact of domain migration.

[0185] Discovery 6: Using non-overlapping users (or items) in learning the preference transfer function greatly helps to improve the recommendation performance. In real-world applications, the overlapping users between two domains are very limited, which hinders the training of the preference transfer function based on this information. This is why SSCDR uses the information of overlapping users and non-overlapping source domain items and LACDR performs better than the EMCDR paradigm (training the mapping function between two domains based only on the information of overlapping users) when using the information of overlapping and non-overlapping users. Note that HCCDR also considers these overlapping users when it learns the transfer function of users using the common features extracted from similar users in the source domain.

[0186] Discovery 7: Personalized preference transfer effectively promotes the performance of CDR tasks. Considering that the user preference relationship between the source domain and the target domain varies from user to user, PTUPCDR and HCCDR learn personalized preference transfer functions for each user, rather than using a universal preference transfer function for all users. The evaluation results listed in Table 3 show that PTUPCDR and HCCDR outperform all other methods, demonstrating the significance of customized transfer functions. At the same time, we can observe from the table that HCCDR outperforms PTUPCDR due to HCCDR's consideration of individual and common features of users when learning personalized preference transfer functions.

[0187] Table 3 Comparison of HCCDR with baselines in three CDR tasks

[0188]

[0189] 4. Ablation Study

[0190] To analyze the effectiveness of the modules proposed in HCCDR, we conduct an ablation study. Specifically, we consider the following four variants of HCCDR:

[0191] HCDDR(GMF) is a variant of HCCDR, where it uses the GMF method to generate the embeddings of users and items, instead of the HLFM component proposed in HCCDR. Note that GMF ignores the user-user and item-item relationships when computing the embeddings.

[0192] HCDDR(N2V) is a variant of HCCDR, which computes the embeddings of users and items using the Node2vec, a homogeneous graph embedding technique, instead of the HLFM component proposed in HCCDR, based on the multi-relational graph we project out.

[0193] HCDDR-CCG is a variant of HCCDR. It removes the Common Feature Generator (CCG) module from the Cluster Enhanced Preference Transfer (CPT) component of HCCDR, which indicates that we remove the common features generated from clustering and only use the individual features of users to train the personalized preference transfer function of users.

[0194] HCDDR(GMF)-CCG is a variant of HCDDR(GMF) where the CCG module is disabled.

[0195] We conduct an ablation study on the three tasks with an 8:2 ratio of training set to test set. Note that similar conclusions can be drawn using other ratio settings. The experimental results are shown in Table 4, and the ablation analysis results are as follows.

[0196] Table 4 Ablation study (best results are shown in bold, and secondary best values are underlined).

[0197]

[0198] Effectiveness of Heterogeneous Latent Factor Modeling (HLFM): We first evaluate the effectiveness of the HLFM component proposed in HCCDR by comparing HCCDR with its two variants, i.e., HCCDR(GMF) and HCCDR(N2V). The results in Table 4 show that the MAE (or RMSE) of HCCDR(GMF) is averagely 1.69% (or 1.64%) higher than that of HCCDR, because the GMF method used in HCCDR(GMF) fails to consider implicit relationships (i.e., user-user relationships and item-item relationships) to compute the embeddings of users and items, but only considers user-item interactions. On the other hand, although HCCDR(N2V) considers all three types of relationships (user-item, user-user, item-item), it applies a homogeneous graph embedding method, i.e., Node2vec, to generate representations for users and items. Since it ignores the heterogeneous nature of these relationships, HCCDR(N2V) produces unsatisfactory recommendation results compared to HCCDR. From another important observation in the table is that HCCDR(N2V) is the best variant, which again demonstrates the importance of learning representations from the three types of relationships. In summary, the two optimization strategies in the HLFM component, i.e., considering all three types of relationships and embedding-oriented heterogeneity, make the proposed HCCDR superior to its two variants.

[0199] Effectiveness of Cluster-enhanced Preference Transfer (CPT): The importance of CPT components can be investigated by comparing HCCDR and its variant HCCDR-CCG. The novelty of CPT is to extract common features from similar users, and use these information together with the individual features of a user to learn a personalized transfer function between two domains. In HCCDR-CCG, the novelty in CPT is eliminated by disabling its key module called Common Characteristics Generator (CCG). The experimental results in Table 4 show that the performance of HCCDR in terms of MAE (or RMSE) is on average improved by 2.43% (or 2.93%) compared to HCCDR-CCG. This demonstrates the effectiveness of learning a personalized transfer bridge between two domains for each user by exploiting common features extracted from similar users.

[0200] Finally, by comparing HCCDR with its variant HCCDR(GMF)-CCG, we find from Table 4 that the MAE (or RMSE) of HCCDR(GMF)-CCG is on average increased by 3.7% (or 3.86%) compared to HCCDR without the advantage brought by HLFM and CPT components, which is a significant growth of error. Such observation again demonstrates the importance of reasonably identifying and handling heterogeneous relationships (i.e., HLFM) when computing the representations of users and items, and the importance of considering both individual and common features of a user (i.e., CPT) when learning a personalized transfer function for a user.

[0201] 5. Visualization Experiments

[0202] To investigate the quality of transferred user embeddings in the target domain, we also compare the HCCDR proposed in this embodiment with the existing models EMCDR and PTUPCDR.

[0203] Specifically, PTUPCDR is the best baseline method, which also learns a personalized transfer function between two domains for each user; EMCDR is the typical paradigm of CDR, which learns a mapping function between source and target domains based only on the information of overlapping users between two domains. Note that these baseline methods, i.e., DCDCSR, DCDIR and PTUPCDR, all follow the framework of EMCDR. We employ the t-SNE tool to visualize the transferred user embeddings of each comparison method, and the ground-truth user embeddings in the target domain are consistently generated by the HLFM component of HCCDR. Due to space limitations, we only discuss the visualization results of Task 1, while the same conclusions can be drawn for Task 2 and Task 3. Note that the ratio of training set to test set is set to 8:2, and we randomly sample 512 user embeddings for each method to visualize clearly. The visualization results are shown in Figure 6 where dark points represent the ground-truth user embeddings in the target domain, and light points represent the transferred user embeddings.Figure 6 The left-to-right scatter plots correspond to the EMCDR, PTUPCDR, and HCCDR models, respectively.

[0204] As shown in Figure 6 (a), the distribution of the transferred user embeddings is significantly different from that of the ground-truth user embeddings, because one of them (i.e., the transferred user embeddings) is scattered, while the other one (i.e., the ground-truth user embeddings) is concentrated in a certain area. This is because the EMCDR learns a universal preference transfer function for all users, which cannot capture the different relationships of user preferences between the source and target domains. In contrast, the sub Figure 6 (b) and 6(c) show that the distribution of the transferred user embeddings is in good agreement with that of the ground-truth user embeddings, because both of these methods customize the preference transfer function for each user.

[0205] At the same time, the sub Figure 6 (c) shows that the proposed HCCDR successfully captures the common characteristics of the user groups in the target domain, because there are many clusters in the distribution of the transferred user embeddings (or the ground-truth user embeddings). In contrast, there are almost no obvious clusters in the sub Figure 6 (b), because the PTUPCDR does not take into account the common characteristics of similar users. This also demonstrates that the proposed clustering-enhanced personalized preference transfer in the HCCDR is quite effective in computing high-quality user embeddings in the target domain.

[0206] Embodiment Two

[0207] This embodiment proposes a cross-domain recommendation system based on the personalized preference transfer model of embodiment one, which includes a computer device programmed or configured to perform the cross-domain recommendation method of the personalized preference transfer model described in embodiment one.

[0208] In summary, the present application optimizes the performance of CDR by addressing two core issues of CDR, i.e., "what to transfer" and "how to transfer". For the first issue, existing methods only utilize explicit user-item relationships to learn user and item embeddings, ignoring implicit user-user and item-item relationships. For the second issue, existing work does not consider user common features derived by analyzing similar users in the process of learning user's personalized preference transfer function between two domains. To this end, we construct a novel clustering-enhanced personalized preference transfer model HCCDR. To address the first issue, the model first models three types of relationships, i.e., user-item, user-user, and item-item, in a multi-relational graph, and then carefully designs a heterogeneous relation-oriented embedding method to learn more effective user and item representations by absorbing the information of the multi-relational graph. To well alleviate the second issue, the model first attempts to combine user common features and individual features to learn user's personalized preference transfer function. Experimental results show that the model significantly outperforms all baselines in two real-world datasets and three CDR tasks. The main contributions are summarized as follows:

[0209] I. To address the limitations of existing CDR methods in "what to transfer" and "how to transfer", we propose a novel heterogeneous and clustering-enhanced personalized preference transfer model (HCCDR) for cross-domain recommendation, which provides better recommendation performance for cross-domain recommendation.

[0210] II. To address the first limitation, we construct a multi-relational graph in each domain to model users and items (user-item relationships, user-user relationships, and item-item relationships), and design an effective heterogeneous relation-oriented embedding method to generate rich representations of users and items based on graph.

[0211] III. To address the second limitation, not only is the personalized preference transfer function of each user trained according to the individual features of the user, but also the common features of the user and similar users are derived by using a soft clustering mechanism for enhancement.

[0212] IV. We conduct extensive experiments on two public datasets. Experimental results show that HCCDR improves the average by 4.94% compared to the best baseline in terms of mean absolute error (MAE) and root mean square error (RMSE).

[0213] The above are only the preferred embodiments of the present application, and do not limit the present application in any form. Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solutions of the present application, and according to the technical essence of the present application, should fall within the scope of protection of the technical solutions of the present application.

Claims

1. A cross-domain recommendation method based on personalized preference transfer model, characterized in that, Comprising the following steps: S1) constructing a personalized preference transfer model, which firstly constructs a multi-relation graph according to the interaction, basic information and corresponding comments of users and items, and calculates the embedding of all users and items in the source domain and the target domain respectively according to the information of the multi-relation graph; Then, according to the learned user representation, a soft clustering method is used to identify similar users for each target user in the source domain, and the individual characteristics of the target user and the common characteristics of the similar users are used to learn the personalized preference transfer function of the target user; Finally, the user embedding transferred from the source domain to the target domain is calculated according to the learned personalized preference transfer function of each user, and the predicted score of the user to the items in the target domain is calculated according to the transferred user embedding to realize the personalized recommendation of the source domain user in the target domain; S2) obtaining the interaction, basic information and corresponding comments of users and items and inputting them into the personalized preference transfer model to obtain the personalized recommendation of each user; In step S1), when the multi-relation graph is constructed according to the interaction, basic information and corresponding comments of users and items, and the embedding of all users and items in the source domain and the target domain is calculated according to the information of the multi-relation graph, it comprises: S101) converting the basic information and user comments related to the user into a user document vector, converting the basic information and user comments related to the item into an item document vector, calculating the similarity probability between each pair of users in the user document vector, and calculating the similarity probability between each pair of items in the item document vector; Then all users and items are summarized as nodes of the multi-relation graph, the corresponding user-user heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of users, the corresponding item-item heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of items, and the corresponding user-item edges in the multi-relation graph are generated according to the historical interaction between users and items. 2.The cross-domain recommendation method based on a personalized preference transfer model according to claim 1, characterized in that, In step S1), when the multi-relation graph is constructed according to the interaction, basic information and corresponding comments of users and items, and the embedding of all users and items in the source domain and the target domain is calculated according to the information of the multi-relation graph, it further comprises: S102) According to the embedding strategy for heterogeneous relationship, the user-user relationship and the item-item relationship in the multi-relationship graph are taken as the similarity relationship, the user-item relationship in the multi-relationship graph is taken as the interaction relationship, and the Euclidean distance is used to measure the distance between two nodes in the latent vector space for each similarity relationship , and the loss function at the minimum is calculated , and all the interaction relationships are modeled as the translation between the nodes in the latent vector space using an explicit translation-based method , and the loss function at the minimum is calculated , finally, the loss function and the loss function are minimized together, and the embedding matrix of the user and the item is obtained. 3.The cross-domain recommendation method based on personalized preference transfer model according to claim 2, characterized in that, In step S101), when the corresponding user-user heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of users, the corresponding item-item heterogeneous edges in the multi-relation graph are generated according to the similarity probability between each pair of items, and the corresponding user-item edges in the multi-relation graph are generated according to the historical interaction between users and items, it comprises: If the similarity probability between a pair of users or the similarity probability between a pair of items is greater than a preset threshold, the corresponding user-user heterogeneous edge or item-item heterogeneous edge in the multi-relation graph is generated, and the similarity probability corresponding to the user-user heterogeneous edge or item-item heterogeneous edge is taken as the weight of the user-user heterogeneous edge or item-item heterogeneous edge; If there is a historical interaction between the user and the item, the corresponding user-item edge in the multi-relation graph is generated, and the quotient of the historical rating of the user to the item and the maximum value in the preset rating matrix is calculated as the weight of the user-item edge. 4.The cross-domain recommendation method based on a personalized preference transfer model according to claim 1, characterized in that, In step S1, according to the learned user representation, similar users are identified for each target user in the source domain using a soft clustering method, and when learning the personalized preference transfer function of the target user using the individual features of the target user and the common features of the similar users of the target user, specifically comprising: S201) group all users in the source domain into K clusters, respectively calculate the assignment probability of assigning the embedding of each user to the kth cluster, and then establish the target distribution to guide the optimization of the unsupervised clustering loss function , and further update the cluster center, and finally multiply the assignment probability of the target user to each cluster by the cluster center of the class, and then accumulate to obtain the common features of the target user; S202) obtaining an interaction item list of the target user in the source domain, using an attention network to calculate the attention score of each item in the interaction item list respectively, and normalizing to obtain the weight of each item, multiplying the embedding of each item in the interaction item list of the target user by the corresponding weight and accumulating to obtain the individual feature of the target user; S203) connecting the common features and individual features of the target user through a connection operation to obtain the transferable features of the target user, using a preset neural network to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user, and then generating the representation of the target user after transfer in the target domain according to the personalized preference transfer function of the target user. 5.The cross-domain recommendation method based on personalized preference transfer model according to claim 4, characterized in that, establishing the target distribution in step S201 to guide the optimization of the unsupervised clustering loss function when further updating the clustering centers, comprising: A soft assignment matrix recording soft assignment information of the source domain of clusters [ 1,..., N ] ∈ N×K , target distribution Y is defined as follows: wherein, ∑ i S i,k is the k-th cluster center o k the soft clustering frequency of S i,k denotes the probability of assigning a user u i to the k-th cluster o k . Utilizing a soft assignment matrix S with a target distribution Y to define a clustering loss function to make the soft assignments close to the target distribution as follows: In each iteration, the set of cluster centers O [ o 1, o 2,..., o K ] are updated by stochastic gradient descent, each cluster center corresponding gradient is computed as follows: where β is expressed as the degrees of freedom of the t-distribution of students, representing the embedding of the user The update of the k-th cluster center is shown as follows: wherein, is a learning rate to control the speed of the cluster center update. 6.The cross-domain recommendation method based on a personalized preference transfer model according to claim 4, characterized in that, The attention score expression of each item in the interaction item list in step S202 is as follows: where the function is an attention network, is a learnable parameter, denotes the embedding of item v j . The weight expression of each item is as follows: wherein, represents the list of interaction items of the user u i in the source domain, v l represents the lth item in the list of interaction items of the user u i in the source domain. 7.The cross-domain recommendation method based on a personalized preference transfer model according to claim 4, characterized in that, In step S203, using a preset neural network to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user, and then generating the representation of the target user after transfer in the target domain according to the personalized preference transfer function of the target user, comprising: The expression of using a preset neural network to learn the personalized preference transfer function of the target user from the source domain to the target domain for the transferable features of the target user is as follows: where the vector ∈ contains parameters of a personalized preference transfer function, function ( ; ) is a two-layer neural network parameterized by parameters, denotes transferrable features of the target user; The vector ∈ reshaped into a matrix to fit the size of the parameters of the preference transfer function, where denotes the embedding dimension of the user; With matrix As parameters, using the target user Personalized preference transfer function to generate target user The transformed representation in the target domain is expressed as follows: wherein, ∈ represents a representation of the target user in the source domain. 8.The cross-domain recommendation method based on a personalized preference transfer model according to claim 4, characterized in that, In step S1, according to the representation of the target user after transfer in the target domain, the predicted score of the target user on each item in the target domain is calculated, specifically comprising: performing dot product operation on the representation of the target user after transfer in the target domain and the representation of each item in the target domain to obtain the predicted score of the target user on each item in the target domain, and generating the personalized recommendation of the target user according to the predicted score of each item. 9.The cross-domain recommendation method based on a personalized preference transfer model according to claim 8, characterized in that, After obtaining the predicted score of the target user on each item in the target domain, it further comprises: training the personalized preference transfer function using a task-based optimization method to minimize the loss between each predicted score and the real score.

10. A cross-domain recommendation system based on personalized preference transfer model, characterized in that, It comprises a computer device programmed or configured to perform the cross-domain recommendation method based on the personalized preference transfer model of any one of claims 1-9.

Images