A knowledge-aware multi-domain recommendation method and system based on hierarchical graph contrastive learning

Abstract

The application provides a knowledge-aware multi-domain recommendation method and system based on hierarchical graph contrast learning, comprising: obtaining a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by weighted summation smoothing target nodes and neighbor nodes, aggregating more preference knowledge based on a hierarchical message passing mechanism through interactive items and connected entities; based on a preset contrast learning method, deleting part of the nodes in the hierarchical nodes for data enhancement of contrast learning; based on the enhanced data, performing cross-multi-domain domain-shared graph contrast learning and specific-domain-only graph contrast learning to identify the intrinsic preference characteristics of a target user; based on the intrinsic preference characteristics of the target user, constructing a multi-task model, optimizing and training the multi-task model to generate a multi-domain recommendation model, and completing multi-domain recommendation through the multi-domain recommendation model. The application solves the problem that the existing multi-domain recommendation is difficult to accurately predict user preferences.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content recommendation, and particularly relates to a knowledge perception multi-domain recommendation method and system based on hierarchical graph contrastive learning. BACKGROUND

[0002] Personalized recommendation systems have been widely used to assist users in making decisions in various recommendation scenarios, such as product recommendation on shopping platforms, short video recommendation, and paper recommendation on knowledge platforms. In order to meet the diverse needs of users, these platforms provide more and more heterogeneous items belonging to different domains, for example, shopping platforms provide products in multiple domains, including books, digital music, clothing, etc. Each user can browse and comment on items in different domains on the platform, and items in each domain can also be interacted by different users. Therefore, in order to better characterize the preferences of the target user, the multi-domain user behavior can be aggregated to understand the diverse characteristics of the target user, thereby improving the recommendation quality in each domain.

[0003] Multi-domain recommendation (MDR) aims to improve the recommendation performance in each domain by utilizing the rich knowledge from multiple domains, which has attracted extensive attention in recent years. The key challenge lies in how to capture the domain-invariant and specific domain preferences of the target user based on the overlapping entities (users or items) across domains. A common solution is based on the multi-task learning strategy, which can regard the recommendation goal in each domain as a task, and then jointly use the multi-domain behavior information to generate the final recommendation. In addition, some other models utilize different techniques (i.e., self-supervised learning, feature clustering, pre-training, and transfer learning) to improve the recommendation performance of users in the target domain. However, despite the extensive research on MDR today, their performance still suffers from some bottleneck problems, and some key challenges have not been well addressed. For example, 1) the small amount of user interaction data leads to inefficient supervision effect. In real-world recommendation systems, because most users only interact with a limited number of items, the cross-dataset from different domains always shows weak supervision problems; in addition, the lack of sample labels also seriously reduces the representation efficiency of users or items in the model training process. Therefore, how to alleviate the sparse interaction problem without a large number of explicit training labels is the key to improving the performance of MDR. 2) The incomplete feature problem makes the existing models pay more attention to the unique domain preferences of the target user in the target domain, ignoring their domain-invariant features in other domains. Simply combining multiple linear training models cannot comprehensively capture the multi-dimensional user preferences, making it difficult to solve the user feature bias problem in the multi-domain scenario.

[0004] Sparse user interaction problem makes it difficult to achieve effective prediction performance in multi-domain recommendation. Although many current technologies (transfer technology, self-supervised learning, etc.) can be applied to the multi-domain recommendation scenario, these existing models still suffer from low-efficiency recommendation results. The feature bias problem causes the prediction accuracy of the current multi-domain recommendation model to be very low. Most recommendation methods only focus on the user features in the target domain, ignoring the domain-invariant preferences of users in other domains, so it is difficult to accurately recommend the user's preferred content. SUMMARY

[0005] The application provides a knowledge-aware multi-domain recommendation method and system based on hierarchical graph contrastive learning to solve the problem that the current multi-domain recommendation is difficult to accurately predict user preferences.

[0006] The application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, comprising:

[0007] Obtain multiple nodes on the hierarchical interaction network, smooth the target node and the neighbor node by weighted summation for hierarchical representation, and aggregate more preference knowledge through the interactive items and their connected entities based on the hierarchical message passing mechanism;

[0008] Based on the preset contrastive learning method, delete part of the nodes in the hierarchical nodes for data enhancement of contrastive learning;

[0009] Based on the enhanced data, identify the intrinsic preference features of the target user through cross-multi-domain domain-shared graph contrastive learning and single-domain graph contrastive learning in a specific domain;

[0010] Based on the intrinsic preference features of the target user, construct a multi-task model, optimize and train the multi-task model to generate a multi-domain recommendation model, and complete multi-domain recommendation through the multi-domain recommendation model.

[0011] According to the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application, the multiple nodes on the hierarchical interaction network are obtained, the target node and the neighbor node are smoothed by weighted summation for hierarchical representation, and more preference knowledge is aggregated through the interactive items and their connected entities based on the hierarchical message passing mechanism, which specifically comprises:

[0012] There are large-scale nodes on the hierarchical interaction network, including target nodes, and each target node has neighbor nodes around it;

[0013] Smooth the features of the target node and its neighbor nodes by weighted summation strategy for hierarchical representation definition;

[0014] Form a higher-level hierarchical representation of users or items by stacking different levels of representation, and aggregate the hierarchical representations obtained on multiple levels to generate the final preference knowledge representation.

[0015] The application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning.

[0016] Through the contrastive learning method, a set proportion of nodes and connection relationships thereof are randomly discarded on each hierarchical view.

[0017] The discarding probability of each node follows the same uniform distribution, and data enhancement is realized.

[0018] The application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning.

[0019] In the domain-shared graph contrastive learning across multiple domains, domain-invariant features of the target user are learned across multiple domains, and the invariant preferences of the user are exhibited by learning the behavior features in all domains.

[0020] The generated different hierarchical views are used to jointly train the contrastive learning task of multiple domains, and domain-invariant features in different hierarchical structures are obtained to identify the intrinsic preference features of the target user.

[0021] The application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning.

[0022] The domain-specific graph contrastive learning in the specific domain includes item-based domain-specific contrastive learning and neighbor-based domain-specific contrastive learning.

[0023] The item-based domain-specific contrastive learning generates positive contrastive examples from the representations of two specific domains of the same interaction node based on the enhanced hierarchical view, and negative contrastive examples can be obtained from the representations of different interaction nodes in each hierarchical view.

[0024] The neighbor-based domain-specific contrastive learning establishes a neighbor-based domain-specific contrastive learning task by using similar neighbor interaction items, can perform high-order connection between nodes in the specific domain, adopts a multi-hop reasoning strategy to obtain high-order preference signals of the target user, and neighbor-guided positive samples and negative samples can provide different preference signals for the proposed model to identify the preference features.

[0025] According to the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided in the application, a multi-task model is constructed based on the internal preference features of a target user, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

[0026] Based on domain sharing and specific domain views, multi-layer knowledge aggregation is performed on users or items, and multi-domain representation is obtained through multiple contrastive learning tasks.

[0027] The unobserved rating value in the target domain is predicted through a multi-domain Bayesian personalized ranking method, and an optimization target is determined.

[0028] Based on the optimization target, a multi-task optimization strategy is executed to train and generate a multi-domain recommendation model for multi-domain recommendation.

[0029] The application also provides a knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning, which comprises:

[0030] A hierarchical representation aggregation module is configured to obtain multiple nodes on a hierarchical interaction network, perform hierarchical representation by weighted summation smoothing target nodes and neighbor nodes, and aggregate more preference knowledge through interactive items and connected entities based on a hierarchical message passing mechanism.

[0031] A data enhancement module is configured to perform data enhancement for contrastive learning by deleting part of the nodes in the hierarchical nodes based on a preset contrastive learning method.

[0032] A graph contrast module is configured to identify the internal preference features of the target user based on enhanced data through cross-multi-domain shared graph contrastive learning and specific-domain unique-domain graph contrastive learning.

[0033] A model generation module is configured to construct a multi-task model based on the internal preference features of the target user, optimize and train the multi-task model to generate a multi-domain recommendation model, and complete multi-domain recommendation through the multi-domain recommendation model.

[0034] The application also provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to any of the above.

[0035] The application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to any of the above.

[0036] The application also provides a computer program product comprising a computer program which, when executed by a processor, implements the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to any one of the above.

[0037] The application provides a knowledge-aware multi-domain recommendation method and system based on hierarchical graph contrastive learning, which realizes hierarchical knowledge representation of users or items through an efficient aggregation model representation strategy based on a knowledge graph, enhances original interaction data and captures more self-supervised semantic signals through a hierarchical node discarding mechanism, and proposes three graph contrastive learning tasks to explore preference features of target users, so as to alleviate the problems of data sparsity and poor multi-domain recommendation prediction performance, and train the proposed multi-domain recommender by using a multi-task learning framework. In addition, compared with various advanced baseline models, the performance improvement of the application on two real-world data sets proves the effectiveness of the proposed method, and verifies that the scheme can realize strong recommendation performance in a sparse interaction scene; not only the domain-shared features of the target user in the multi-domain can be utilized, but also the unique-domain features of the target user in a specific domain can be aggregated. In addition, the multi-domain recommendation model proposed by the application not only realizes good prediction performance in terms of ablation test and hierarchical knowledge aggregation depth, but also exhibits better recommendation accuracy in online testing. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0039] Figure 1 is one of the process flow diagrams of the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application;

[0040] Figure 2 is the second process flow diagram of the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application;

[0041] Figure 3 is the third process flow diagram of the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application;

[0042] Figure 4 is the fourth process flow diagram of the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application;

[0043] Figure 5 is the fifth process flow diagram of the knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning provided by the application;

[0044] Figure 6 Fig. 1 is a schematic diagram of a module connection of a knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning provided by the present application;

[0045] Figure 7 Fig. 2 is an architecture diagram of a knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning provided by the present application;

[0046] Figure 8 Fig. 3 is a structural schematic diagram of an electronic device provided by the present application.

[0047] Reference signs:

[0048] 110: hierarchical representation aggregation module; 120: data enhancement module; 130: graph contrastive module; 140: model generation module;

[0049] 810: processor; 820: communication interface; 830: memory; 840: communication bus. DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The present application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, which comprises the following steps: Figures 1-5 The present application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, which comprises the following steps:

[0052] S100, a plurality of nodes on a hierarchical interaction network are acquired, hierarchical representation is performed on a target node and neighbor nodes by weighted summation smoothing, more preference knowledge is aggregated through the interaction of items and the entities connected thereto based on a hierarchical message passing mechanism;

[0053] S200, based on a preset contrastive learning method, data enhancement is performed on the contrastive learning of part of the nodes in the hierarchical nodes;

[0054] S300, based on the enhanced data, cross-multi-domain shared graph contrastive learning in a specific domain and unique domain graph contrastive learning in a specific domain are performed to identify the intrinsic preference features of a target user;

[0055] S400, a multi-task model is constructed based on the intrinsic preference features of the target user, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

[0056] The present application is inspired by graph contrastive learning and knowledge graph technology, and combines knowledge-aware graph context information into a self-supervised training model to improve the performance of multi-domain recommendation. Specifically, based on the defined hierarchical interaction network, the present application first uses a LightGCN aggregation model to obtain the hierarchical representation of users and items on the knowledge graph. Then, the present application proposes a hierarchical node dropout strategy to enhance the original interaction samples and capture more self-supervised signals. In particular, the present application designs a novel graph contrastive network to learn the preference features of the target user. In the proposed graph contrastive learning network, the present application designs a domain-shared contrastive learning task to obtain user domain-invariant features across different domains; at the same time, based on similar neighbors and clicked items in the target domain, the present application designs two contrastive learning tasks to obtain specific domain features of the target user. Finally, the present application proposes a multi-domain Bayesian personalized ranking mechanism to learn the contrastive loss of the model in combination with a multi-task learning architecture.

[0057] Generally, given multiple domain interaction matrices, knowledge graphs and hierarchical interaction networks, the present application aims to build a novel knowledge-aware multi-domain recommendation model based on hierarchical graph contrastive learning to predict potential items that the target user may like. The present application mainly includes the following four parts: hierarchical knowledge representation and aggregation; data enhancement on multi-domain hierarchical views; graph contrastive network on multi-domain and specific domain; multi-task training model.

[0058] Obtain multiple nodes on the hierarchical interaction network, smooth the target node and the neighbor node by weighted summation, and aggregate more preference knowledge based on the hierarchical message passing mechanism, specifically including:

[0059] S101, there are large-scale nodes on the hierarchical interaction network, including target nodes, each target node has neighbor nodes around it;

[0060] S102, smooth the features of the target node and its neighbor nodes by a weighted summation strategy, and define hierarchical representation;

[0061] S103, form a higher level hierarchical representation of users or items by stacking different level representations, and aggregate the hierarchical representations obtained on multiple levels to generate a final preference knowledge representation.

[0062] In the present application, based on the powerful representation ability of graph neural networks, a novel hierarchical knowledge representation mechanism is first designed to aggregate the preference knowledge of the target user in each domain. In order to process and learn the representation of large-scale nodes on the hierarchical interaction network and KG, the present application uses the efficient aggregation principle of the LightGCN aggregation model, and uses a simple weighted summation strategy to smooth the features of the target node and its neighbor nodes. Therefore, the hierarchical representation of users and items in each domain is defined as follows:

[0063]

[0064]

[0065] Where h represents the h-th layer in the hierarchical interaction network, and Let u and i represent the propagation representations of user u and item i in layer h, respectively. N represents the set of all items that interact with user u in the h-th layer. i (h) This represents the set of all users who interact with i in the h-th layer.

[0066] To fully leverage the high-order relationships between multiple nodes, this invention designs a hierarchical messaging mechanism on the knowledge graph, and then aggregates more preference knowledge through interacting items and their connected entities. To this end, this invention stacks different hierarchical representations to form higher-level hierarchical representations of users or items. Finally, this invention aggregates the hierarchical representations obtained at multiple levels to generate the final representation.

[0067]

[0068]

[0069] in, This represents the hierarchical importance of the h-th level representation in the final representation. For This invention can be obtained using the same computational method as described above. Therefore, based on the hierarchical message passing mechanism, the final representation of the target node aggregates semantic signals from multiple domains, thus providing a more comprehensive characterization of the target user's preference features. Clearly, current embedding methods (TransE, MF, etc.) fail to capture more supervisory signals from sparse user-item interactions across multiple domains, thus hindering their ability to learn better unbiased representations. In contrast, the hierarchical representation solution proposed in this invention can provide more preference signals for model training.

[0070] Based on a pre-defined contrastive learning method, data augmentation is performed by deleting some nodes in the hierarchical nodes and then performing contrastive learning. Specifically, this includes:

[0071] S201. Using a comparative learning method, a set proportion of nodes and their connection relationships are randomly discarded on each layered view.

[0072] S202. The drop probability of each node follows the same uniform distribution, thus achieving data augmentation.

[0073] In the present application, inspired by the rapid development of data augmentation techniques in graph contrastive learning, the present application attempts to combine self-supervised signals into node representation learning. However, data augmentation operations cannot be simply applied directly to the recommendation system, while the behavior preferences of the target user's multi-hop neighbors can reveal their potential intentions to some extent. Therefore, the present application uses the neighbor information of the target user in the knowledge graph to better learn the representation of the node.

[0074] In the proposed framework, the present application designs to generate different hierarchical views derived from multi-domain graph mapping. Therefore, based on the advanced GraphCL method, the present application proposes a hierarchical node deletion strategy as a data augmentation method for contrastive learning in multi-domain recommendation models. Specifically, given an interaction graph G, the present application randomly discards a certain proportion of nodes and their associated connections on each hierarchical view, and the principle behind this operation is that discarding some nodes will not affect the semantic degree of the hierarchical network, and the discarding probability of each node follows the same uniform distribution. To this end, the present application defines the data augmentation operation on the user-item interaction graph as follows:

[0075]

[0076]

[0077]

[0078] where h represents different hierarchical structures, k is a random sampling number, represents the inner product of vectors, and are two mask vectors, which depend on the user-item interaction graph on the hierarchical structure h. Through the above mask operation, the present application can pay more attention to the intrinsic preference features of the target node, and enhance the robustness of entity representation learning to noisy connections. Therefore, based on the two random data augmentation operations of each domain, the enhanced hierarchical views and In addition, it can also be seen from the above data augmentation operation that the hierarchical representation of entities in the knowledge graph stores multi-hop semantic signals from neighbors in different hierarchical structures.

[0079] Based on the enhanced data, the present application identifies the intrinsic preference features of the target user through cross-multi-domain shared graph contrastive learning and single-domain graph contrastive learning in specific domains, which specifically includes:

[0080] In cross-multi-domain shared graph contrastive learning, the domain-invariant features of the target user are learned across multiple domains, and the invariant preferences of the user are exhibited by learning the behavior features in all domains.

[0081] ​By jointly training the multi-domain contrastive learning task by using the generated different hierarchical views, domain-invariant features in different hierarchical structures are obtained to identify the intrinsic preference features of the target user.

[0082] Based on the enhanced data, the intrinsic preference features of the target user are identified by cross-multi-domain domain-shared graph contrastive learning and specific-domain individual-domain graph contrastive learning, and the method further comprises:

[0083] S301, the individual-domain graph contrastive learning in the specific domain comprises an item-based individual-domain contrastive learning and a neighbor-based individual-domain contrastive learning;

[0084] S302, the item-based individual-domain contrastive learning, based on the item-based individual-domain contrastive learning task, for each item node, positive samples are generated from the representations of two specific domains of the same interaction node based on the enhanced hierarchical view, and negative samples can be obtained from the representations of different interaction nodes in each hierarchical view;

[0085] S303, the neighbor-based individual-domain contrastive learning, by using similar neighbor interaction items to establish the neighbor-based individual-domain contrastive learning task, high-order connections between nodes in the specific domain can be performed, a multi-hop inference strategy is adopted to obtain high-order preference signals of the target user, neighbor-guided positive samples and negative samples can provide different preference signals for the proposed model to identify the preference features.

[0086] In the application, graph contrastive learning is a very promising self-supervised training method, which is based on contrastive loss in the feature space, and the data features are enhanced by maximizing the consistency between two views on the same graph. Unlike previous graph contrastive models, the application combines some semantic knowledge provided by the knowledge graph into the graph contrastive learning framework to enhance the self-supervised signal of the model. Based on the different hierarchical views obtained in each field, the application designs three graph contrastive learning tasks to verify the effectiveness of the proposed method. Among them, the domain-shared contrastive learning task is proposed to learn the shared features of the target user across all domains, and the specific-domain graph contrastive learning is proposed to learn the individual-domain features of the target user in each specific domain.

[0087] The challenge of domain-shared contrastive learning is to learn the domain-invariant features of the target user across multiple domains, where the invariant preferences of the user are revealed by learning their behavior features in all domains. In order to guide the acquisition of multi-domain knowledge-aware preference features, the application proposes a hierarchical semantic invariant method based on KG representation technology. Therefore, the hierarchical semantic invariant coefficient Hsi pq is defined as follows:

[0088] Hsi pq =s(f p (x p ,ζ p φ(ep ,r p i p ))+f q (x q ,ζ q φ(e q ,r q i q ))) (8)

[0089] Where p and q represent different hierarchical structures, f p The embedding aggregation function represents knowledge awareness, and s represents the cosine function used to evaluate x. p and x q Consistency between them, ζ p and ζ q The hierarchical factor is represented. Based on the hierarchical semantic invariant coefficients, the hierarchical shared representation of users and projects is modeled through an MLP framework.

[0090]

[0091]

[0092] Where σ is an ELU nonlinear function, p,q∈H. Then, the proposed method uses generated hierarchical views to jointly train a multi-domain contrastive learning task. Specifically, for any node's hierarchical shared representation... Considered as part of the same batch of S Positive samples. On the other hand, the present invention randomly samples from other batches S′. To pair negative instances. Based on the InfoNCE algorithm, the domain-shared contrastive learning loss across multiple domains is defined as follows:

[0093]

[0094] Where sim(·) is a cosine function that measures the similarity between positive and negative samples, and τ is a temperature parameter. By minimizing the domain-shared contrast loss, this invention obtains domain-invariant features in different hierarchical structures to identify the inherent preference features of target users.

[0095] In addition to exploring domain-shared features of the target user across multiple domains, this invention further mines additional feature information from different domains as complementary collaborative signals to jointly predict the target user's final preferences. Therefore, based on the hierarchical views obtained in each domain, this invention attempts to perform domain-specific contrastive losses on both views for mutual supervision. To this end, based on the target user's neighbors and the items they interact with, this invention designs two domain-specific contrastive learning tasks.

[0096] Project-based intra-domain contrastive learning. In each hierarchical structure, those items that the target user interacted with reveal their current latent intention, which plays a key role in generating the final preference feature. Therefore, the present invention designs a project-based intra-domain contrastive learning task based on the above idea. For each item node, the present invention can obtain the representations of two specific domains from the same interaction node based on the enhanced hierarchical view Positive pair examples are generated. Similarly, negative pair examples can be obtained from the representations of different interaction nodes in each hierarchical view. Therefore, the project-based intra-domain contrastive loss of the present invention is defined as follows:

[0097]

[0098] where I u is the set of items that the target user interacts with in a specific domain, and τ is the temperature parameter. The project-based intra-domain contrastive loss is calculated on all positive and negative pair examples in a small batch, so that they reach a consensus on the latent semantics.

[0099] Neighbor-based intra-domain contrastive learning. Although the clicked items can directly provide explicit preference signals for the target user, the newly online recommendation system often has only a small amount of interaction data, resulting in a serious data sparsity problem of the system. Therefore, based on the principle that "similar users tend to like similar items", the behavior features of those similar neighbors in a specific domain can also provide implicit preference signals for the model. To this end, the present invention assumes that the interaction items of similar neighbors can be used to propose a neighbor-based intra-domain contrastive learning task, which can perform high-order connections between nodes in a specific domain. If N u is the set of similar neighbors of the target user u in a specific domain, then the neighbor-based intra-domain contrastive loss is defined as follows:

[0100]

[0101] where, is the representation of the target node, and are the representations of the neighbor nodes and non-neighbor nodes in a specific domain, respectively. In order to explore the semantic preferences of more neighbors in a specific domain, the present invention adopts a multi-hop reasoning strategy to obtain high-order preference signals of the target user. In addition, the positive and negative samples guided by these neighbors can provide different preference signals for the proposed model, so that the neighbor-based and clicked item-based intra-domain contrastive learning tasks can complement each other and collaboratively train the objective function.

[0102] Intuitively, the present invention combines these two intra-domain contrastive losses to form the final specific domain contrastive loss function.

[0103] L Specific = LIspecific +L NSpecific (14)

[0104] The multi-task model is constructed based on the inherent preference characteristics of the target user, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model, specifically comprising:

[0105] S401, based on domain sharing and specific domain view, after multi-layer knowledge aggregation is performed on users or items, multi-domain representation is obtained through multiple contrast learning tasks;

[0106] S402, the unobserved rating value in the target domain is predicted through a multi-domain Bayesian personalized ranking method, and an optimization target is determined;

[0107] S403, the multi-domain recommendation model is trained based on the optimization target by executing a multi-task optimization strategy, and multi-domain recommendation is performed.

[0108] In the application, based on domain sharing and specific domain view, after multi-layer knowledge aggregation is performed on users or items, the application obtains their multi-domain representation through three contrast learning tasks. The final predicted score is generated by the inner product of the user and item representation.

[0109]

[0110] Wherein, and are the final multi-domain representation of users and items respectively.

[0111] Based on the obtained self-supervised signal, the application executes a multi-task optimization strategy to train the proposed multi-domain recommendation model. Inspired by the original BPR and its variant CBPR, the application proposes a multi-domain Bayesian personalized ranking method to predict the unobserved rating value in the target domain. The prediction formula is as follows:

[0112]

[0113] Wherein, H represents the total number of hierarchical structures, λ h represents the regularization parameter of each hierarchical structure, σ(·) is the sigmoid function, O t and O h are training samples from the target domain and other domains. Therefore, the optimization target of the application is defined as follows.

[0114]

[0115] Wherein, Θ is the learnable model parameter, λ1 and λ2 represent the contrast parameters to determine the self-supervised signal strength of the shared level and the specific level, and λ3 represents the regularization parameter of the joint loss function.

[0116] The application provides a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, realizes hierarchical knowledge representation of users or items through an efficient aggregation model representation strategy based on a knowledge graph, enhances original interaction data and captures more self-supervised semantic signals through a hierarchical node discarding mechanism, and proposes three graph contrastive learning tasks to explore preference features of target users, so as to alleviate the problems of data sparseness and poor multi-domain recommendation prediction performance, and train the proposed multi-domain recommender by using a multi-task learning framework. In addition, compared with various advanced baseline models, the performance improvement of the application on two real-world data sets proves the effectiveness of the proposed method, and verifies that the scheme can realize strong recommendation performance in a sparse interaction scene; not only the domain-shared features of the target user in the multi-domain can be utilized, but also the unique-domain features of the target user in the specific domain can be aggregated. In addition, the multi-domain recommendation model proposed by the application not only realizes good prediction performance in terms of ablation test and hierarchical knowledge aggregation depth, but also exhibits better recommendation accuracy in online testing.

[0117] Reference Figure 6 and Figure 7 The application also discloses a knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning, and the system comprises:

[0118] The hierarchical representation aggregation module 110 is used for acquiring a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by smoothing a target node and neighbor nodes through weighted summation, and aggregating more preference knowledge through an interaction item and connected entities based on a hierarchical message passing mechanism;

[0119] The data enhancement module 120 is used for deleting part of nodes in the hierarchical nodes based on a preset contrastive learning method to perform data enhancement of contrastive learning;

[0120] The graph contrastive module 130 is used for identifying intrinsic preference features of a target user by performing domain-shared graph contrastive learning in cross multi-domain and unique-domain graph contrastive learning in a specific domain based on the enhanced data;

[0121] The model generation module 140 is used for constructing a multi-task model based on the intrinsic preference features of the target user, optimizing and training the multi-task model to generate a multi-domain recommendation model, and completing multi-domain recommendation through the multi-domain recommendation model.

[0122] The hierarchical representation aggregation module 110 is used for acquiring a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by smoothing a target node and neighbor nodes through weighted summation, and aggregating more preference knowledge through an interaction item and connected entities based on a hierarchical message passing mechanism;

[0123] The features of the target node and the neighbor nodes are smoothed through a weighted summation strategy to perform hierarchical representation definition;

[0124] By stacking different hierarchical representations to form a higher-level hierarchical representation of the user or item, the aggregated hierarchical representations obtained at multiple levels generate a final preference knowledge representation.

[0125] The data enhancement module 120 randomly discards a set proportion of nodes and their connection relationships on each hierarchical view through a contrast learning method.

[0126] The discard probability of each node follows the same uniform distribution, achieving data enhancement.

[0127] The graph contrast module 130 learns domain-invariant features of the target user across multiple domains in cross-multi-domain shared graph contrast learning, and exhibits the invariant preferences of the user by learning behavior features in all domains.

[0128] By jointly training the contrast learning task of multiple domains using the generated different hierarchical views, domain-invariant features in different hierarchical structures are obtained to identify the intrinsic preference features of the target user.

[0129] The domain-specific graph contrast learning includes item-based domain-specific contrast learning and neighbor-based domain-specific contrast learning.

[0130] The item-based domain-specific contrast learning generates positive contrast examples from the representations of two specific domains of the same interaction node based on the enhanced hierarchical view, and negative contrast examples can be obtained from the representations of different interaction nodes in each hierarchical view.

[0131] The neighbor-based domain-specific contrast learning establishes a neighbor-based domain-specific contrast learning task using similar neighbor interaction items, enabling high-order connections between nodes in a specific domain, and using a multi-hop inference strategy to obtain high-order preference signals of the target user. Neighbor-guided positive and negative samples can provide different preference signals for the proposed model to identify preference features.

[0132] The model generation module 140 obtains multi-domain representations through multiple contrast learning tasks based on domain sharing and specific domain views after performing multi-layer knowledge aggregation on users or items.

[0133] The unobserved rating values in the target domain are predicted through a multi-domain Bayesian personalized ranking method to determine the optimization objective.

[0134] Based on the optimization objective, a multi-task optimization strategy is executed to train a multi-domain recommendation model for multi-domain recommendation.

[0135] The application provides a knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning, realizes hierarchical knowledge representation of users or items through an efficient aggregation model representation strategy based on a knowledge graph, enhances original interaction data and captures more self-supervised semantic signals through a hierarchical node discarding mechanism, and proposes three graph contrastive learning tasks to explore preference features of a target user, so as to alleviate the problems of data sparseness and poor multi-domain recommendation prediction performance, and a multi-task learning framework is used to train the proposed multi-domain recommender. In addition, compared with various advanced baseline models, the performance improvement of the application on two real-world data sets proves the effectiveness of the proposed method, and verifies that the scheme can realize strong recommendation performance in a sparse interaction scene; not only the domain-shared features of the target user in the multi-domain can be utilized, but also the unique-domain features of the target user in a specific domain can be aggregated. In addition, the multi-domain recommendation model proposed by the application not only realizes good prediction performance in terms of ablation test and hierarchical knowledge aggregation depth, but also exhibits better recommendation accuracy in online testing.

[0136] Figure 8 An entity structure diagram of an electronic device is shown as follows, Figure 8 The electronic device can include a processor 810, a communications interface 820, a memory 830 and a communications bus 840, wherein the processor 810, the communications interface 820 and the memory 830 complete mutual communication through the communications bus 840. The processor 810 can invoke a logical instruction in the memory 830 to execute a knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, which comprises the following steps: obtaining a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by weighted summation smoothing target nodes and neighbor nodes, and aggregating more preference knowledge based on a hierarchical message passing mechanism through an interactive item and an entity connected thereto.

[0137] Based on the preset contrastive learning method, part of the nodes in the hierarchical nodes are deleted for data enhancement of contrastive learning;

[0138] Based on the enhanced data, domain-shared graph contrastive learning in cross multi-domain and unique-domain graph contrastive learning in a specific domain are performed to identify intrinsic preference features of a target user;

[0139] Based on the intrinsic preference features of the target user, a multi-task model is constructed, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

[0140] Further, the logic instructions in the memory 830 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or parts of the present application that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0141] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute a knowledge-aware multi-domain recommendation method based on hierarchical graph contrast learning provided by the above-mentioned methods, the method comprising: obtaining a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by smoothing target nodes and neighbor nodes through weighted summation, and aggregating more preference knowledge through interactive items and connected entities based on a hierarchical message passing mechanism;

[0142] Based on the preset contrast learning method, part of the nodes in the hierarchical nodes are deleted for data enhancement of contrast learning;

[0143] Based on the enhanced data, the intrinsic preference features of the target user are identified through cross-multi-domain shared graph contrast learning and single-domain graph contrast learning in a specific domain;

[0144] Based on the intrinsic preference features of the target user, a multi-task model is constructed, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

[0145] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, the computer program is executed by a processor to implement a knowledge-aware multi-domain recommendation method based on hierarchical graph contrast learning provided by the above-mentioned methods, the method comprising: obtaining a plurality of nodes on a hierarchical interaction network, performing hierarchical representation by smoothing target nodes and neighbor nodes through weighted summation, and aggregating more preference knowledge through interactive items and connected entities based on a hierarchical message passing mechanism;

[0146] Based on the preset contrast learning method, part of the nodes in the hierarchical node are deleted for data enhancement of contrast learning;

[0147] Based on the enhanced data, the intrinsic preference features of the target user are identified through cross-domain shared graph contrast learning and specific domain individual graph contrast learning in multiple domains.

[0148] Based on the intrinsic preference features of the target user, a multi-task model is constructed, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

[0149] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0150] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0151] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning, characterized in that, The method comprises the following steps: Obtaining a plurality of nodes on a hierarchical interaction network, smoothing target nodes and neighbor nodes by weighted summation, and aggregating more preference knowledge through the interaction of items and their connected entities based on a hierarchical message passing mechanism; Based on the preset contrast learning method, part of the nodes in the hierarchical nodes are deleted for data enhancement of contrast learning; Based on the enhanced data, the intrinsic preference features of the target user are identified through cross-domain shared graph contrast learning and single-domain graph contrast learning in a specific domain; Based on the intrinsic preference features of the target user, a multi-task model is constructed, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model.

2. The knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to claim 1, characterized in that, The method comprises the following steps: There are a large number of nodes on the hierarchical interaction network, including target nodes, and each target node has neighbor nodes around it; The features of the target nodes and their neighbor nodes are smoothed by a weighted summation strategy to define a hierarchical representation; Higher-level hierarchical representations of users or items are formed by stacking different levels of representation, and the final preference knowledge representation is generated by aggregating the hierarchical representations obtained at multiple levels.

3. The knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to claim 1, characterized in that, The method comprises the following steps: Through the contrast learning method, a certain proportion of nodes and their connection relationships are randomly discarded on each hierarchical view; The discarding probability of each node follows the same uniform distribution, achieving data enhancement.

4. The knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to claim 1, characterized in that, The method comprises the following steps: In the cross-domain shared graph contrast learning, the domain-invariant features of the target user are learned across multiple domains, and the user's invariant preferences are exhibited by learning the behavior features in all domains; By using the generated different hierarchical views to jointly train the multi-domain contrast learning task, the domain-invariant features in different hierarchical structures are obtained to identify the intrinsic preference features of the target user.

5. The knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to claim 4, characterized in that, The method comprises the following steps: In the single-domain graph contrast learning in a specific domain, it includes single-domain contrast learning based on items and single-domain contrast learning based on neighbors; The single-domain contrast learning based on items generates positive samples based on enhanced hierarchical views from the representations of two specific domains of the same interaction node for each item node, and negative samples can be obtained from the representations of different interaction nodes in each hierarchical view; The neighbor-based private domain contrastive learning utilizes the interaction items of similar neighbors to establish a neighbor-based private domain contrastive learning task, can perform high-order connections between nodes in a specific domain, adopts a multi-hop reasoning strategy to obtain high-order preference signals of the target user, and neighbor-guided positive and negative samples can provide different preference signals for the proposed model and identify preference features.

6. The knowledge-aware multi-domain recommendation method based on hierarchical graph contrastive learning according to claim 1, characterized in that, The multi-task model is constructed based on the internal preference features of the target user, the multi-task model is optimized and trained to generate a multi-domain recommendation model, and multi-domain recommendation is completed through the multi-domain recommendation model. Specifically, the method comprises the following steps: Based on domain sharing and specific domain view, after performing multi-layer knowledge aggregation on users or items, multi-domain representation is obtained through multiple contrastive learning tasks; An unobserved rating value in the target domain is predicted through a multi-domain Bayesian personalized ranking method to determine an optimization target; Based on the optimization target, a multi-task optimization strategy is executed to train and generate a multi-domain recommendation model for multi-domain recommendation.

7. A knowledge-aware multi-domain recommendation system based on hierarchical graph contrastive learning, characterized in that, The system comprises: A hierarchical representation aggregation module is configured to obtain multiple nodes on a hierarchical interaction network, perform hierarchical representation by weighted summation smoothing target nodes and neighbor nodes, and aggregate more preference knowledge through the interaction items and their connected entities based on a hierarchical message passing mechanism; A data enhancement module is configured to delete part of the nodes in the hierarchical nodes based on a preset contrastive learning method to perform data enhancement for contrastive learning; A graph contrast module is configured to identify the internal preference features of the target user based on the enhanced data through domain-shared graph contrast learning in cross-multi-domain and private domain graph contrast learning in specific domains; A model generation module is configured to construct a multi-task model based on the internal preference features of the target user, optimize and train the multi-task model to generate a multi-domain recommendation model, and complete multi-domain recommendation through the multi-domain recommendation model.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the knowledge-aware multi-domain recommendation method based on hierarchical graph contrast learning according to any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the knowledge-aware multi-domain recommendation method based on hierarchical graph contrast learning according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the knowledge-aware multi-domain recommendation method based on hierarchical graph contrast learning according to any one of claims 1 to 6.

Images