A collaborative communication method and device based on feature disentanglement

Abstract

The present invention discloses a collaborative communication method and device based on feature disentanglement. The method includes: adaptively updating the original heterogeneous graph based on the feature similarity of users and products from the perspectives of domain invariance and domain specificity, generating an adaptive updated heterogeneous graph; utilizing graph convolution technology to aggregate node features on the original heterogeneous graph and the adaptive updated heterogeneous graph, and weightedly fusing the aggregated node features to obtain user domain invariant features, domain specific features, and product features; utilizing a meta-network to personalized transfer the user domain invariant features aggregated from the source domain to the target domain, and weightedly fusing them with the user domain invariant features of the target domain to obtain the final user domain invariant features of the target domain; utilizing two constraint losses based on contrastive learning to achieve efficient disentanglement of dual-domain user features; and summing the two interaction prediction scores as the final interaction prediction score to obtain the matching relationship between the user and the product. The device includes: a processor and a memory.

Description

Technical Field

[0001] The present invention relates to the field of cross-domain collaborative recommendation, and in particular to a collaborative communication method and device based on feature disentanglement. Background Art

[0002] Since the recommendation system can effectively capture the personalized preferences of users, it has been widely used in many consumer platforms such as Taobao and Amazon. However, with the rapid increase in the number of users and products, the interaction data between users and products has become very sparse, which makes collaborative filtering-based [1] (CF) and matrix factorization [2] Traditional recommendation methods based on MF technology suffer from data sparsity, making it difficult to accurately model user preferences. To alleviate this data sparsity, cross-domain recommendation has emerged. Its goal is to transfer user preference knowledge from other domains to the target domain, enabling the model to generate a more complete representation of user preferences in the target domain, thereby improving recommendation performance in the target domain.

[0003] Existing cross-domain recommendation technologies can be roughly divided into two categories: one is hybrid cross-domain recommendation and the other is disentangled cross-domain recommendation. Hybrid cross-domain recommendation aims to fuse user features of the source domain and the target domain in different combinations. Hu et al. [3] proposed a cross-network based on fully connected layers to perform feature fusion. Liu et al. [4] A bidirectional feature transfer module is designed in the graph convolution process to perform weighted feature fusion. But in fact, not all user preferences learned in the source domain can be transferred to the target domain. If the current source domain and target domain are movies and books respectively, the user preferences learned in the movie domain may include: movie plots, movie shooting styles, etc., but for the book domain, only movie plots can help the book domain to recommend books. The movie shooting style not only cannot help the book recommendation, but may interfere with the book recommendation. The above-mentioned hybrid cross-domain recommendation ignores this problem, migrates features that should not be migrated, and is prone to negative transfer problems, so only suboptimal solutions can be achieved. In order to avoid the negative transfer problem, the disentangled cross-domain recommendation method divides user preferences into domain-specific features and domain-invariant features, where domain-specific features only work within the domain, while domain-invariant features work in both domains. How to achieve effective feature disentanglement is an urgent problem that needs to be solved by the disentangled cross-domain recommendation method. Zhang et al. [5] Orthogonal loss is used to achieve disentanglement of intra-domain features. Cao et al. [6] The variational inference framework is introduced to achieve disentanglement by increasing the KL divergence distance between domain-invariant features and domain-specific features. Although existing disentanglement-based cross-domain recommendation methods have achieved good results, they do not achieve complete feature disentanglement and have the following three main shortcomings:

[0004] First, the subtle differences in interactive behaviors are ignored during the generation of disentangled representations. Each interactive behavior has its underlying intention, which means that different interactive behaviors contribute differently to the generation of different disentangled representations. For example, when intending to transfer a user's preferences in the clothing domain to the sports domain, the behavior of purchasing a wool skirt due to its wool material is more helpful in generating domain-specific features, because wool material has little relevance in the sports domain. However, existing research ignores the subtle differences in interactive behaviors and assumes that each behavior contributes equally to the generation of different disentangled representations, which may lead to incomplete feature disentanglement.

[0005] Second, existing user feature disentanglement processes are unrelated to the features of the recommended items. Whether using orthogonal loss or KL divergence, their disentanglement processes are solely related to the user features themselves. However, the model ultimately needs to consider both user and item features to make recommendation predictions. This means that considering disentanglement solely from the perspective of user features while ignoring item features will only achieve suboptimal disentanglement performance.

[0006] Third, it ignores the personalized transfer of user domain invariant features. After obtaining domain invariant representation, the existing disentangled cross-domain recommendation methods usually use weighted fusion. [9] or fully connected operation [5] However, this strategy ignores the personalized differences in user preferences and cannot efficiently utilize the learned domain-invariant features.

[0007] Therefore, given the shortcomings of the above-mentioned existing solutions in cross-domain recommendation, an efficient method is urgently needed to address them. Summary of the Invention

[0008] This invention provides a collaborative communication method and apparatus based on feature disentanglement. The invention determines the contribution of different interactions to the generation of different disentangled representations based on the similarity between the user's domain-invariant features and domain-specific features and the corresponding product features. It utilizes a meta-network to generate a personalized transfer function to achieve personalized cross-domain transfer of domain-invariant features, thereby simultaneously improving dual-domain recommendation performance. It also utilizes two contrastive learning-based constraint losses to achieve efficient and thorough feature disentanglement, as described below.

[0009] In a first aspect, a collaborative communication method based on feature disentanglement is provided, the method comprising:

[0010] From the perspectives of domain invariance and domain specificity, the original heterogeneous graph is adaptively updated according to the feature similarity of users and products to generate an adaptive updated heterogeneous graph.

[0011] Graph convolution technology is used to aggregate node features on the original heterogeneous graph and the adaptively updated heterogeneous graph, and the aggregated node features are weighted fused to obtain user domain invariant features, domain-specific features, and product features.

[0012] The meta-network is used to transfer the user-domain invariant features aggregated from the source domain to the target domain in a personalized manner, and then weightedly fuse them with the user-domain invariant features of the target domain to obtain the final user-domain invariant features of the target domain.

[0013] We utilize two contrastive learning-based constraint losses to achieve efficient disentanglement of dual-domain user features.

[0014] The interaction prediction scores of user domain invariant features and product features, and the interaction prediction scores of user domain specific features and product features are calculated respectively. The two interaction prediction scores are summed as the final interaction prediction score to obtain the matching relationship between users and products.

[0015] The original heterogeneous graph is adaptively updated according to the feature similarity between the user and the product to generate an adaptive updated heterogeneous graph:

[0016] Similarity filter score:

[0017]

[0018] Among them, c∈{I,S}, I, S represent domain-invariant features and domain-specific features respectively, s(·,·) is the dot product operation, e u,l and e i,l are the features of user u and item i in the lth convolutional layer, * can represent domain A or domain B, and σ(·) is a nonlinear activation function;

[0019] Adaptively update the adjacency matrix:

[0020]

[0021] in, Represents the normalized adjacency matrix.

[0022] Furthermore, the formula for weighted fusion is as follows:

[0023]

[0024] Among them, α is a hyperparameter, and They are user features and product features after information aggregation through the original heterogeneous graph at layer l+1, and They are respectively the user features and product features after information aggregation through adaptively updating the heterogeneous graph at the l+1 layer.

[0025] Among them, the final target domain invariant features are:

[0026]

[0027] Among them, β is a hyperparameter, is the domain-invariant feature of domain A, e A,T (u,i) is the transfer feature from domain B to domain A;

[0028]

[0029] in, is the personalized transfer function of user u, is the personalized transfer function of product i, is the domain-invariant feature of user u in domain B;

[0030]

[0031] in, and are the personalized transfer function matrices of users and products respectively, and They represent the personalized extractors on the user side and the product side respectively, which are composed of fully connected layers. and They represent user meta-knowledge and product meta-knowledge respectively.

[0032] The constraint loss includes similarity contrast learning loss and exclusive contrast learning loss;

[0033] The similarity contrastive learning loss is:

[0034]

[0035] Among them, φ(·,·) is the cosine similarity to measure the similarity of representations, τ p is the temperature coefficient of contrast loss, A collection of common users. is the domain-invariant feature of user u, is the migration feature of user u;

[0036] The exclusive contrastive learning loss can be expressed as:

[0037]

[0038] in, is the interaction behavior observed in domain A, s(·,·) is the point multiplication operation to measure the mutual information between user features and product features, L represents the number of graph convolution layers, is the domain-specific feature of user i in domain A, is the characteristic of product i in domain A, is the domain-specific feature of user u after the l-th layer of graph convolution in domain B.

[0039] In the second aspect, a collaborative communication device based on feature disentanglement is provided, the device comprising: a processor and a memory, wherein the memory stores program instructions, and the processor calls the program instructions stored in the memory to enable the device to execute any one of the methods described in the first aspect.

[0040] According to a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any one of the methods described in the first aspect.

[0041] The beneficial effects of the technical solution provided by the present invention are:

[0042] 1. This invention targets the field of cross-domain collaborative recommendation and disentangles user features into domain-specific features and domain-invariant features. It also transfers only the domain-invariant features, effectively avoiding the negative transfer problem.

[0043] 2. This paper introduces adaptive graph update technology, which can mine the subtle differences in each interaction behavior in generating different features, adaptively update the original heterogeneous graph, and thus generate more fine-grained disentanglement features;

[0044] 3. The present invention uses the meta-network to learn effective personalized transfer functions for users, thereby achieving personalized migration of invariant features in the user domain and meeting the personalized differences between different users;

[0045] 4. The present invention designs two disentanglement constraint losses based on contrastive learning. On the one hand, they can narrow the distance between the domain-invariant features of the same user in two domains. On the other hand, they can ensure that the domain-specific features of the user within the domain do not contain domain-invariant information, effectively achieving the purpose of feature disentanglement. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of a collaborative communication method based on feature disentanglement. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0048] In order to solve the shortcomings of the above cross-domain recommendation method, the embodiment of the present invention proposes a collaborative propagation method and device based on feature disentanglement. [6]The proposed method and algorithm introduce the idea of ​​adaptive graph updating. Based on the similarity of user and product features, it can mine the subtle differences of different user-product interactions in generating different disentangled features, and adaptively update the original heterogeneous graph to generate more fine-grained disentangled features. As an effective self-supervisory technology, contrastive learning has been used to alleviate the data sparsity problem in recommendation systems due to its ability to narrow the distance between different views of the same node and at the same time widen the distance between different nodes. Xie et al. [7] Achieving consistency of cross-domain user features using contrastive learning. Guo et al. [8] Contrastive learning is used to narrow the distance between the features of the same user in different views. In view of the successful application of contrastive learning in recommendation systems, the proposed algorithm designs two constraint losses based on contrastive learning to achieve efficient feature disentanglement. Different from Xie et al. [7] The proposed method,the similarity contrast loss in the proposed algorithm only narrows the distance between the,domain-invariant pairs without involving domain-specific features,,which is more in line with actual needs.,Different from Zhang et al. [5] The proposed method, the exclusive contrast loss in the proposed algorithm uses the mutual information between users and products to separate the domain-invariant information contained in domain-specific features to achieve efficient disentanglement. In addition, considering the differences between users, that is, some users may pay more attention to domain-invariant factors such as price when purchasing products, while some users may pay more attention to domain-specific factors such as brand, which is different from Zhu et al. [9] The proposed method, an embodiment of the present invention utilizes a meta-network to generate a personalized transfer function for each user, thereby achieving an effective domain-invariant feature migration process. The proposed method has been verified on the widely used Amazon dataset and its effectiveness in cross-domain recommendation has been verified through multiple experiments.

[0049] Example 1

[0050] A collaborative communication method based on feature disentanglement, see Figure 1 , the method comprises the following steps:

[0051] 101: Construct the original heterogeneous user-item graphs based on the implicit feedback of the source and target domains respectively;

[0052] 102: Randomly initialize user and product features, and project the randomly initialized user features into domain-invariant initial features and domain-specific initial features;

[0053] 103: Adaptively update the original heterogeneous graph based on the feature similarity of users and products from both domain-invariant and domain-specific perspectives to generate an adaptively updated heterogeneous graph;

[0054] 104: Use graph convolution technology to aggregate node features on the original heterogeneous graph and the adaptively updated heterogeneous graph, and perform weighted fusion on the aggregated node features to obtain user domain invariant features, domain-specific features, and product features;

[0055] 105: Using the meta-network, the user-domain invariant features aggregated from the source domain are personalized and transferred to the target domain, and weightedly fused with the user-domain invariant features of the target domain to obtain the final user-domain invariant features of the target domain;

[0056] 106: Efficient disentanglement of dual-domain user features using two constrained losses based on contrastive learning;

[0057] 107: Calculate the interaction prediction score between the user domain invariant feature and the product feature, and the interaction prediction score between the user domain specific feature and the product feature respectively, and sum the two interaction prediction scores as the final interaction prediction score to obtain the matching relationship between the user and the product.

[0058] Among them, the interactive prediction score of the user domain invariant feature and the product feature, and the interactive prediction score of the user domain specific feature and the product feature are special technical terms in this field and will not be described in detail in the embodiment of the present invention.

[0059] In summary, the embodiment of the present invention can, through the above steps 101-107, explore the subtle differences in generating different features for each interactive behavior, adaptively update the original heterogeneous graph, and thus generate more fine-grained disentangled features; realize personalized migration of user domain invariant features to meet the personalized differences between different users; and can narrow the distance between the invariant features of the same user's two domains, so that the domain-specific features of the user within the domain do not contain domain-invariant information, effectively achieving the purpose of feature disentanglement.

[0060] Example 2

[0061] The solution in Example 1 is further introduced below with reference to specific calculation formulas and examples.

[0062] Step 201: Constructing respective user-item heterogeneous graphs based on implicit feedback from the source domain and the target domain;

[0063] The steps are: first obtain the common user set of the source domain and the target domain (denoted as domain A and domain B) Dual-domain product collection and The embodiment of the present invention can achieve simultaneous improvement of dual-domain performance. Therefore, the source domain can refer to domain A, and the target domain is domain B. However, the source domain can also refer to domain B, and the target domain is domain A. This will not be described in detail below. And the interaction matrix of the dual domains is established based on the interaction records between users and products. and If user u and product i have interacted, then The value of is 1. If user u and product i have never interacted, then The value of is 0, where According to the interaction matrix obtained, the original heterogeneous graph of user-item relationship is established. Each original heterogeneous graph contains user and item nodes in their respective domains. ui When the value of is 1, it means that the user node u and the product node i are connected, otherwise, there is no connection.

[0064] Step 202: Randomly initialize user and product features, and project the randomly initialized user features into domain-invariant initial features and domain-specific initial features;

[0065] In this embodiment, the embodiment of the present invention uses Xavier initialization to randomly initialize user and product features, and obtains the A domain user feature matrix A domain product feature matrix B-domain user feature matrix B domain product feature matrix The present invention proposes the need to disentangle user features into domain-invariant features and domain-specific features. Domain-specific features only aid prediction within their own domain, while domain-invariant features enable dual-domain prediction. To achieve this goal, the present invention applies two independent projection mappings to randomly initialized user features within each domain, thereby forming a user domain-invariant initial feature matrix and a domain-specific initial feature matrix. The specific mapping formulas are as follows:

[0066]

[0067] Among them, * can represent domain A or domain B. is the user domain invariant initial feature matrix, is the user domain specific initial feature matrix, is the randomly initialized user feature matrix, W i * and W s * is a learnable mapping matrix, and is the learnable bias, σ(·) is the nonlinear activation function, and ⊙ is the element-wise product.

[0068] The two independent projection mappings described above can prevent the mutual influence between user domain invariant features and domain-specific features during the subsequent graph convolution process. In this embodiment of the present invention, we only focus on disentanglement of user features. Therefore, we do not involve projection mapping for product features.

[0069] Step 203: Adaptively update the original heterogeneous graph based on the feature similarity between users and products from both domain-invariant and domain-specific perspectives to generate an adaptively updated heterogeneous graph;

[0070] In an embodiment of the present invention, in order to generate more fine-grained disentangled features, a parameter-free adaptive graph update module is proposed. The design purpose of this module is to explore the subtle differences in the generation of different features for each interactive behavior. In order to avoid the waste of computing resources and network update interference caused by introducing more training parameters, this module is calculated only based on existing user features and product features, without involving new parameters. Generally speaking, when considering the clothing field and the book field, if user u purchases clothing i due to factors in a specific field, such as cotton material (this factor has a low correlation with the book field), then it can be inferred that in the clothing field, compared with domain-invariant features, this purchase behavior contributes more to the generation of domain-specific features. Therefore, the similarity of each user feature and product feature is calculated from the domain-invariant and domain-specific perspectives to measure the subtle differences in the current interactive behavior in generating different features. The calculation formula of the similarity filter score is as follows:

[0071]

[0072] Among them, c∈{I,S}, I, S represent domain-invariant features and domain-specific features respectively, s(·,·) is the dot product operation, e u,l and e i,l are the features of user u and item i in the lth convolutional layer respectively.

[0073] After obtaining the similarity filter scores of all interactions, they are combined with the original normalized adjacency matrix to generate the final adaptively updated adjacency matrix:

[0074]

[0075] in, represents the normalized adjacency matrix, which is composed of the interaction matrix It is obtained through Laplace regularization.

[0076] Based on the above steps, we can explore the subtle differences in each interaction behavior in generating different features, and adaptively update the original heterogeneous graph based on the similarity between user and product features, thereby generating an adaptive updated heterogeneous graph.

[0077] Step 204: Utilize graph convolution technology to aggregate node features on the original heterogeneous graph and the adaptively updated heterogeneous graph, and perform weighted fusion on the aggregated node features to obtain domain-invariant features, domain-specific features, and product features of the user.

[0078] In practical applications, graph convolution technology has been widely used in recommendation systems due to its ability to aggregate high-level knowledge. In the embodiments of the present invention, graph convolution technology is also used to aggregate high-level information representations of users and products. Specifically, information aggregation is performed on the original heterogeneous graph and the adaptively updated graph, and then the final user and product features of each layer are obtained through weighted fusion. Finally, the final features of all convolutional layers are averaged and pooled to obtain the final user and product features after information aggregation.

[0079] First, in each layer of graph convolution, information aggregation is performed on the original heterogeneous graph to obtain the original information aggregation features:

[0080]

[0081] in, is the original normalized adjacency matrix Perform a transpose operation, and They are respectively the user features and product features after information aggregation through the original heterogeneous graph at the l+1 layer.

[0082] Afterwards, information aggregation is performed on the adaptively updated heterogeneous graph to obtain the adaptively updated information aggregation features:

[0083]

[0084] in, and They are respectively the user features and product features after information aggregation through adaptively updating the heterogeneous graph at the l+1 layer.

[0085] Then, after obtaining the original information aggregation features and the adaptively updated information aggregation features, the final user and product aggregation features of each layer of graph convolution are obtained through weighted fusion. The weighted fusion formula is as follows:

[0086]

[0087] Here, α is a hyperparameter.

[0088] Finally, after obtaining the final features of L convolutional layers, average pooling is performed to obtain the final user and product features after information aggregation:

[0089]

[0090] In this embodiment of the present invention, the detangling operation of the product features is not considered. Therefore, for the final product features, this embodiment of the present invention fuses the domain-invariant product features and the domain-specific product features:

[0091]

[0092] Where f(·,·) is the average pooling operation.

[0093] Step 205: Using the meta-network, the user-domain invariant features aggregated from the source domain are personalized and transferred to the target domain, and weightedly fused with the user-domain invariant features of the target domain to obtain the final user-domain invariant features of the target domain.

[0094] Simply weighting and fusing or concatenating the user-domain invariant features of the two domains will ignore user individual differences, resulting in only suboptimal cross-domain migration. Therefore, embodiments of the present invention design a meta-network-based personalized transfer module to achieve personalized cross-domain transfer of domain-invariant features. Specifically, a user-side and product-side personalized extractor consisting of fully connected layers is constructed, generating a personalized transfer function for the user. This personalized transfer function is then used to achieve personalized cross-domain transfer of the source domain user-domain invariant features, which are then weighted and fused with the target domain invariant features.

[0095] First, two independent fully connected layers are used as the user-side personalized extractor and the product-side personalized extractor, respectively. User meta-knowledge, representing the user context encoding information, is input into the user-side personalized extractor to generate the user's personalized transfer matrix. Product meta-knowledge, representing the product context encoding information, is input into the product-side personalized extractor to generate the product's personalized transfer function matrix. Taking the transfer of domain B knowledge to domain A as an example, domain B is the source domain and domain A is the target domain. The formula for extracting the personalized transfer function matrix is ​​as follows:

[0096]

[0097] in, and are the personalized transfer function matrices of users and products respectively, and They represent the personalized extractors on the user side and the product side respectively, which are composed of fully connected layers. and Represent user meta-knowledge and product meta-knowledge respectively. The user and product meta-knowledge are represented as:

[0098]

[0099] Among them, || represents the full connection operation, and Represent the neighbor nodes of user u and product i respectively, is the user domain invariant feature matrix of domain A, is the user domain invariant feature matrix of domain B, is the user domain-specific feature matrix of domain A, is the product feature matrix of domain A, is the characteristic of product i in domain A, and The domain-specific features and domain-invariant features of domain user u are respectively A.

[0100] After obtaining the personalized transfer function, the personalized cross-domain transfer is realized by using the personalized transfer function to transfer the invariant features of the source domain and user domain. Taking the transfer of domain B knowledge to domain A as an example, the transfer process formula is as follows:

[0101]

[0102] in, is the personalized transfer function of user u, is the personalized transfer function of product i, is the domain-invariant feature of user u in domain B.

[0103] Finally, the transferred features transferred from the source domain to the target domain are weightedly fused with the domain-invariant features of the target domain to obtain the final target domain domain-invariant features. The formula is as follows:

[0104]

[0105] Among them, β is a hyperparameter, is the domain-invariant feature of domain A, e A,T (u,i) is the transfer feature from domain B to domain A, ew ,F (u,i) is the final domain-invariant feature of domain A.

[0106] Step 206: Utilize two contrastive learning based constraint losses to achieve efficient disentanglement of dual-domain user features;

[0107] Contrastive learning, with its ability to reduce the distance between different views of the same node while increasing the distance between different nodes, has been used to alleviate data sparsity in recommendation systems. In this embodiment of the present invention, two loss constraints based on contrastive learning are proposed to help the model achieve efficient disentanglement. Specifically, the similarity contrastive loss is used to reduce the distance between domain-invariant features of the same user in two domains. The exclusive contrastive loss is used to exclude domain-invariant information contained in domain-specific features of users within a domain.

[0108] Since user domain invariant features work simultaneously in both domains, for the same user, their domain invariant features in both domains should be similar. To achieve this goal, the present invention proposes a contrastive loss based on InfoNCE to narrow the distance between the domain invariant features of other domains and the domain invariant features of the current domain. Specifically, the transfer features of the same user and the current domain features are used as positive examples, and the transfer features of different users and the current domain features are used as negative examples. The similarity contrastive learning loss is calculated as follows:

[0109]

[0110] Among them, φ(·,·) is the cosine similarity to measure the similarity of representations, τ p is the temperature coefficient of contrast loss, A collection of common users. is the domain-invariant feature of user u, is the migration feature of user u.

[0111] For ideal feature disentanglement, domain-specific features should not contain domain-invariant information that can enable cross-domain prediction. Therefore, the embodiment of the present invention combines mutual information with contrastive learning to exclude domain-invariant information contained in domain-specific features. The specific operation is to ensure that the mutual information between domain-specific features of other domains and product features of this domain is less than the mutual information between domain-specific features of this domain and product features of this domain. Taking domain A as an example, the exclusive contrastive learning loss can be expressed as:

[0112]

[0113] in, is the interaction behavior observed in domain A, s(·,·) is the point multiplication operation to measure the mutual information between user features and product features, L represents the number of graph convolution layers, is the domain-specific feature of user u in domain A, is the characteristic of product i in domain A, is the domain-specific feature of user u after the l-th layer of graph convolution in domain B.

[0114] Through the above two loss constraints based on contrastive learning, the consistency of dual-domain invariant features and the separation of domain-invariant features and domain-specific features within a domain can be achieved, thereby achieving the goal of ideal disentanglement.

[0115] Step 207: Calculate the final interactive prediction score of the user domain invariant features, domain specific features, and product features to obtain the matching relationship between the user and the product;

[0116] In this embodiment of the present invention, the final matching relationship loss selects BPR loss, which is a classic pairing loss. This loss encourages the model to give higher prediction scores to existing interactions:

[0117]

[0118] in, is the training set, i + is the product that user u has interacted with, i - is a product that user u has not interacted with, σ(·) is a nonlinear activation function, and the prediction score of user u for product i is calculated as follows:

[0119]

[0120] Among them, i includes the above i + and i - , is the final domain-invariant feature of user u, is the domain-specific feature of user u, is the characteristic of commodity i.

[0121] Finally, considering all the loss terms comprehensively, we can get the final overall loss function of the network:

[0122]

[0123] in, is the BPR recommendation loss, and They are similarity contrast loss and exclusive contrast loss, is L2 regularization, λ p ,λ e ,λ l The hyperparameters are set respectively.

[0124] Example 3

[0125] The feasibility of the methods in Example 1 and Example 2 will be verified below with reference to Tables 1 to 5. In this embodiment, the present invention uses four widely used Amazon datasets, specifically Sport&Phone, Sport&Cloth, Elec&Phone, and Elec&Cloth, as described below for details:

[0126] The Amazon dataset records user reviews of products on the Amazon website and is a classic dataset for recommendation systems. It covers 42 fields, each containing information such as user profiles, user ratings, user reviews, and product descriptions. In this example, only user rating information is retained. Based on previous experience, two related fields are selected, such as sports and mobile phones. Users and products with fewer than five interaction records are first filtered out, and only user records shared by both fields are retained as the final dataset. Details of the final four Amazon datasets are shown in Table 1:

[0127] Table 1 Information of four Amazon cross-domain datasets

[0128]

[0129] Evaluation metrics include: Hit Rate (HR@K) (whether the target product is in the top K recommendation lists) and Normalized Discounted Cumulative Gain (NDCG@K) (the position of the target product in the top K recommendation lists). In this embodiment, K is set to 10, and each experiment is randomly run five times. The average of the five experimental results is reported to ensure credibility.

[0130] The present invention is implemented using Python 3.8, PyTorch 1.11.0, and CUDA 11.3. All experiments are conducted on an NVIDIA 3090 GPU and trained using the Adam optimization algorithm. Specific hyperparameter settings are shown in Table 2. In addition, for different datasets, the user and product feature embedding dimensions are 128, α and β are selected from {0.01, 0.05, 0.1, 0.2, 0.3, 0.5, 0.7}, and λ is 0. e Choose from {0.005, 0.01, 0.05, 0.1, 0.2, 0.3, 0.5}, λ p Choose among {0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1}.

[0131] Table 2 Hyperparameter settings

[0132] parameter Setting Values Learning rate 0.001 Batch size 1024 Dropout 0.5 <![CDATA[τ p ]]> 0.05 <![CDATA[λ l ]]> 0.05 L 5 Epoch 100

[0133] Comparison Algorithms: In this experiment, this method is compared with 12 recommendation methods, of which 4 are single-domain methods and 8 are cross-domain methods. Among them, the single-domain methods include: Bayesian personalized ranking algorithm (BPR)

[10] ,Neural collaborative filtering (NCF)

[11] ,Lightweight graph convolution (Simplifying and powering graph convolution network for recommendation, LightGCN)

[12] Disentangled contrastive collaborative filtering (DCCF)

[13] Cross-domain methods include: Collaborative cross networks for cross-domain recommendation (CoNet) [3] ,Deep dual transfer cross domain recommendation (DDTCDR)

[14] ,Dual metric learning for effective and efficient cross-domain recommendation (DML)

[15] ,Multi-level attention-based domain disentanglement for BCDR (MADD) for bidirectional cross-domain recommendation

[16] ,Toward equivalent transformation of user preferences in cross-domain recommendation,ETL

[17] ,Learning disentangled representations for cross-domain recommendation (DisenCDR) [6] ,Disentangled Contrastive Learning for cross-domain recommendation (DCCDR) [5] and Cross domain recommendation via bi-directional transfer graph collaborative filtering networks (BITGCF) [4] .

[0134] Table 3 Comparative experimental results with advanced baseline methods on Sport&Phone and Sport&Cloth datasets (%)

[0135]

[0136] Table 4 Comparative experimental results with advanced baseline methods on Elec&Phone and Elec&Cloth datasets (%)

[0137]

[0138] Tables 3 and 4 show the comparison results of the embodiments of the present invention with existing advanced baseline methods on four Amazon datasets. The best algorithm performance is bolded and the runner-up algorithm performance is underlined, where the p-value is calculated by the proposed algorithm and the runner-up algorithm. The smaller the p-value, the higher the confidence. In the experiment, first, compared with the single-domain recommendation method, the cross-domain recommendation method showed better recommendation performance overall, indicating that migrating useful user knowledge across domains can improve the recommendation performance of the model. Secondly, whether it is a single-domain recommendation method or a cross-domain recommendation method, the overall performance of the recommendation method based on graph convolution (such as LightGCN, DCCF, DCCDR, BITGCF, etc.) is better than that of the recommendation method that does not use graph convolution. In addition, among the cross-domain recommendation methods, the disentangled cross-domain recommendation method (DCCDR, DisenCDR) achieved satisfactory recommendation performance. Finally, regardless of the dataset, the present method (GDCCDR) showed the highest recommendation performance in all indicators. The superior results of the proposed method can be attributed to the following three points:

[0139] 1. The proposed method can exploit the subtle differences in different user-item interaction behaviors to generate different disentanglement features, adaptively update the original heterogeneous graph, and thus generate a more fine-grained disentanglement representation, ultimately improving the overall recommendation performance.

[0140] 2. The proposed method utilizes loss constraints based on contrastive learning to achieve the effective disentanglement goal.

[0141] 3. The proposed method uses the meta-network to realize personalized migration of domain-invariant features, which can meet the preference differences of different users, thereby making the most of cross-domain transfer features and ultimately achieving the optimal solution.

[0142] To further validate the effectiveness of the proposed method, this embodiment of the present invention conducted ablation experiments on the main components of the proposed method on two datasets to evaluate the impact of each module on algorithm performance. Table 5 reports the performance results after ablation of each module. From Table 5, we can draw the following observations.

[0143] Table 5 Ablation experiment results (%)

[0144]

[0145] The performance comparison between the first and second rows shows that removing the exclusive contrastive loss causes a decrease in the model's recommendation performance. Experimental results show that combining contrastive learning with mutual information can effectively exclude domain-invariant information contained in user-domain-specific features, thereby achieving effective disentanglement and ultimately improving recommendation performance. The performance comparison between the first and third rows shows that removing the similarity contrastive loss causes a decrease in the model's recommendation performance. Experimental results show that using contrastive learning to narrow the distance between domain-invariant features across domains for the same user can help the model achieve effective de-entanglement, ultimately improving recommendation performance. The performance comparison between the first and fourth rows shows that removing the adaptive heterogeneous graph update technique causes a decrease in the model's recommendation performance. Experimental results show that exploiting the subtle differences in disentangled representations generated by different user-item interaction pairs can help the model generate more fine-grained disentangled representations, ultimately improving recommendation performance. The performance comparison between the first and fifth rows shows that removing the personalized transfer of domain-invariant features causes a decrease in the model's recommendation performance. Experimental results show that the meta-network-based personalized transfer module can meet the preference differences between different users, thereby making the most of cross-domain invariant features and ultimately improving the recommendation performance.

[0146] The above experiments have verified that the collaborative propagation method based on feature disentanglement proposed in the embodiment of the present invention can effectively disentangle user features into domain-invariant features and domain-specific features, and perform personalized transfer of domain-invariant features, thereby alleviating the data sparsity problem in the recommendation system and ultimately achieving improved dual-domain recommendation performance. The proposed algorithm innovatively utilizes the meta-network to perform personalized cross-domain transfer of domain-invariant features based on users and business methods, thereby achieving more effective knowledge transfer. The proposed algorithm innovatively introduces contrastive learning into cross-domain recommendation and proposes two constraint losses based on contrastive learning to achieve efficient disentanglement.

[0147] Example 4

[0148] A collaborative communication device based on feature disentanglement, the device comprising: a processor and a memory, wherein the memory stores program instructions, and the processor calls the program instructions stored in the memory to cause the device to perform the following method steps in Example 1:

[0149] From the perspectives of domain invariance and domain specificity, the original heterogeneous graph is adaptively updated according to the feature similarity of users and products to generate an adaptive updated heterogeneous graph.

[0150] Graph convolution technology is used to aggregate node features on the original heterogeneous graph and the adaptively updated heterogeneous graph, and the aggregated node features are weighted fused to obtain user domain invariant features, domain-specific features, and product features.

[0151] The meta-network is used to transfer the user-domain invariant features aggregated from the source domain to the target domain in a personalized manner, and then weightedly fuse them with the user-domain invariant features of the target domain to obtain the final user-domain invariant features of the target domain.

[0152] We utilize two contrastive learning-based constraint losses to achieve efficient disentanglement of dual-domain user features.

[0153] The interaction prediction scores of user domain invariant features and product features, and the interaction prediction scores of user domain specific features and product features are calculated respectively. The two interaction prediction scores are summed as the final interaction prediction score to obtain the matching relationship between users and products.

[0154] Among them, the original heterogeneous graph is adaptively updated according to the feature similarity between users and products, and the adaptive updated heterogeneous graph is generated as follows:

[0155] Similarity filter score:

[0156]

[0157] Among them, c∈{I,S}, I, S represent domain-invariant features and domain-specific features respectively, s(·,·) is the dot product operation, e u,l and e i,l are the features of user u and item i in the I-th convolutional layer, * can represent domain A or domain B, and σ(·) is a nonlinear activation function;

[0158] Adaptively update the adjacency matrix:

[0159]

[0160] in, Represents the normalized adjacency matrix.

[0161] Among them, the formula of weighted fusion is as follows:

[0162]

[0163] Among them, α is a hyperparameter, and They are user features and product features after information aggregation through the original heterogeneous graph at layer l+1, and They are respectively the user features and product features after information aggregation through adaptively updating the heterogeneous graph at the l+1 layer.

[0164] Furthermore, the final target domain invariant features are:

[0165]

[0166] Among them, β is a hyperparameter, is the domain-invariant feature of domain A, e A,T (u,i) is the transfer feature from domain B to domain A;

[0167]

[0168] in, is the personalized transfer function of user u, is the personalized transfer function of product i, is the domain-invariant feature of user u in domain B;

[0169]

[0170] in, and are the personalized transfer function matrices of users and products respectively, and They represent the personalized extractors on the user side and the product side respectively, which are composed of fully connected layers. and They represent user meta-knowledge and product meta-knowledge respectively.

[0171] Furthermore, the constraint loss includes similarity contrastive learning loss and exclusive contrastive learning loss;

[0172] The similarity contrastive learning loss is:

[0173]

[0174] Among them, φ(·,·) is the cosine similarity to measure the similarity of representations, τ p is the temperature coefficient of contrast loss, A collection of common users. is the domain-invariant feature of user u, is the migration feature of user u;

[0175] The exclusive contrastive learning loss can be expressed as:

[0176]

[0177] in, is the interaction behavior observed in domain A, s(·,·) is the point multiplication operation to measure the mutual information between user features and product features, L represents the number of graph convolution layers, is the domain-specific feature of user u in domain A, is the characteristic of product i in domain A, is the domain-specific feature of user u after the l-th layer of graph convolution in domain B.

[0178] It should be noted here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0179] The execution subjects of the above-mentioned processor 1 and memory 2 can be computers, single-chip microcomputers, microcontrollers and other devices with computing functions. In specific implementation, the embodiment of the present invention does not limit the execution subject and it is selected according to the needs of actual application.

[0180] Data signals are transmitted between the memory 2 and the processor 1 via the bus 3 , which will not be described in detail in the embodiment of the present invention.

[0181] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored program, and when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiment.

[0182] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0183] It should be noted here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0184] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0185] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted via computer-readable storage media. Computer-readable storage media can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that integrates one or more available media. Available media can include magnetic media or semiconductor media, etc.

[0186] References

[0187] [1]Badrul Sarwar, George Karypis, Joseph Konstan and John Riedl: Item-based collaborative filtering recommendation algorithms. WWW, 2001: 285-295.

[0188] [2] Xiangnan He, Hanwang Zhang, Min-Yen Kan and Tat-Seng Chua: FastMatrix Factorization for Online Recommendation with Implicit Feedback. SIGIR, 2016: 549-558.

[0189] [3]Guangneng Hu, Yu Zhang and Qiang Yang: Conet: Collaborative crossnetworks for cross-domain recommendation. CIKM, 2018: 667-676.

[0190] [4]Meng Liu, Jianjun Li, Guohui Li and Peng Pan: Cross domain recommendation via bi-directional transfer graph collaborative filtering networks. CIKM, 2020: 885-894.

[0191] [5]Ruohan Zhang,Tianzi Zang,Yanmin Zhu,Chunyang Wang,Ke Wang andJiadi Yu:Disentangled Contrastive Learning for Cross-Domain Recommendation.DASFAA,2023:163-178.

[0192] [6] Jiangxia Cao, Xixun Lin, Xin Cao, Jing Ya, Tingwen Liu and Bin Wang: Disencdr: Learning disentangled representations for cross-domain recommendation. SIGIR, 2022: 267-277.

[0193] [7]Ruobing Xie,Qi Liu,Liangdong Wang,Shukai,Bo Zhang and Leyu Lin:Contrastive cross-domain recommendation in matching.KDD,2022:4226-4236.

[0194] [8] Xiaobo Guo, Shaoshuai Li, Naicheng Guo, Jiangxia Cao, Xiaolei Liu, Qiongxu Ma, Runsheng Gan and Yunan Zhao: Disentangled Representations Learning for Multi-Target Cross-Domain Recommendation. ACM Transactions on Information Systems, 2023, 41(4): 1-27. [9] Feng Zhu, Chaochao Chen, Yan Wang, Guanfeng Liu and Xiaolin Zheng: Dtcdr: A framework for dual-target cross-domain recommendation. CIKM, 2019: 1533-1542.

[0195]

[10] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner and LarsSchmidt-Thieme: BPR: Bayesian personalized ranking from implicit feedback. UAI, 2012.

[0196]

[11] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu and Tat-seng Chua: Neural collaborative filtering. WWW, 2017: 173-182.

[0197]

[12] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang and MengWang: Lightgcn: Simplifying and powering graph convolution network for recommendation. SIGIR, 2020: 639-648.

[0198]

[13] Xubin Ren,Lianghao Xia,Jiashu Zhao,Dawei Yin and Chao Huang:DisentangledContrastive Collaborative Filtering.SIGIR,2023.

[0199]

[14] Pan Li and Alexander Tuzhilin:DDTCDR:Deep dual transfer crossdomainrecommendation.WSDM,2020:331-339.

[0200]

[15] Pan Li and Alexander Tuzhilin:Dual Metric Learning for Effectiveand EfficientCross-Domain Recommendations.IEEE Transactions on Knowledge andData Engineering,2023,35(1):321-334.

[0201]

[16] Xinyue Zhang,Jingjing Li,Hongzu Su,Lei Zhu and Hengtao Shen:Multi-level Attention-based Domain Disentanglement for BCDR.ACM Transactions on Information Systems,2023,41(4):1-24.

[0202]

[17] Xu Chen, Ya Zhang, Ivor W Tsang, Yuangang Pan and Jingchao Su: TowardEquivalent Transformation of User Preferences in Cross DomainRecommendation. ACM Transactions on Information Systems, 2023, 41(1):1-31.

[0203] Unless otherwise specified, the embodiments of the present invention do not limit the models of the components. Any component that can perform the above functions may be used.

[0204] Those skilled in the art will understand that the accompanying drawings are only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0205] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A collaborative communication method based on feature disentanglement, characterized in that: The method comprises: From the perspectives of domain invariance and domain specificity, the original heterogeneous graph is adaptively updated according to the feature similarity of users and products to generate an adaptive updated heterogeneous graph. Graph convolution technology is used to aggregate node features on the original heterogeneous graph and the adaptively updated heterogeneous graph, and the aggregated node features are weighted fused to obtain user domain invariant features, domain-specific features, and product features. The meta-network is used to transfer the user-domain invariant features aggregated from the source domain to the target domain in a personalized manner, and then weightedly fuse them with the user-domain invariant features of the target domain to obtain the final user-domain invariant features of the target domain. We utilize two contrastive learning-based constraint losses to achieve efficient disentanglement of dual-domain user features. The interaction prediction scores of user domain invariant features and product features, and the interaction prediction scores of user domain specific features and product features are calculated respectively. The two interaction prediction scores are summed as the final interaction prediction score to obtain the matching relationship between users and products.

2. The collaborative communication method based on feature disentanglement according to claim 1, characterized in that: The original heterogeneous graph is adaptively updated according to the feature similarity between users and products to generate an adaptive updated heterogeneous graph: Similarity filter score: Among them, c∈{I,S}, I, S represent domain-invariant features and domain-specific features respectively, s(·,·) is the dot product operation, e u,l and e i,l are the features of user u and item i in the lth convolutional layer, * represents domain A or domain B, and σ(·) is a nonlinear activation function; Adaptively update the adjacency matrix: in, Represents the normalized adjacency matrix.

3. The collaborative communication method based on feature disentanglement according to claim 1, characterized in that: The formula for weighted fusion is as follows: Among them, α is a hyperparameter, and They are user features and product features after information aggregation through the original heterogeneous graph at layer l+1, and They are respectively the user features and product features after information aggregation through adaptively updating the heterogeneous graph at the l+1 layer.

4. The collaborative communication method based on feature disentanglement according to claim 1, characterized in that: The final target domain invariant feature is: Among them, β is a hyperparameter, is the domain-invariant feature of domain A, e A,T (u,i) is the transfer feature from domain B to domain A; in, is the personalized transfer function of user u, is the personalized transfer function of product i, is the domain-invariant feature of user u in domain B; in, and are the personalized transfer function matrices of users and products respectively, and They represent the personalized extractors on the user side and the product side respectively, which are composed of fully connected layers. and They represent user meta-knowledge and product meta-knowledge respectively.

5. The collaborative communication method based on feature disentanglement according to claim 1, characterized in that: The constraint loss includes similarity contrastive learning loss and exclusive contrastive learning loss: The similarity contrastive learning loss is: Among them, φ(·,·) is the cosine similarity to measure the similarity of representations, τ p is the temperature coefficient of contrast loss, A collection of common users. is the domain-invariant feature of user u, is the migration feature of user u; The exclusive contrastive learning loss is: in, is the interaction behavior observed in domain A, s(·,·) is the point multiplication operation to measure the mutual information between user features and product features, L represents the number of graph convolution layers, is the domain-specific feature of user u in domain A, is the characteristic of product i in domain A, is the domain-specific feature of user u after the l-th layer of graph convolution in domain B.

6. A collaborative communication device based on feature disentanglement, characterized in that: The device includes: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 5.

Images