A knowledge graph recommendation method based on multi-view contrast learning

Abstract

The application belongs to the technical field of recommendation system, and provides a knowledge graph recommendation method based on multi-view contrast learning. Four different views are comprehensively considered and constructed from different perspectives, including a collaborative knowledge graph composed of a knowledge graph and a user-item interaction graph, a user-item interaction graph, and a user-user graph and a project-project graph constructed based on the user-item interaction graph. In addition, the intersection and union ratio is applied to reasonably construct the user-user graph and the project-project graph, and a receptive field is designed to avoid introducing more noise. Then, the four views are subjected to contrast learning at the local and global levels, aiming to mine the collaborative information between users and projects, between users, and between projects, and the global structure information in a self-supervised manner, thereby alleviating the problem of sparse supervision signals.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of recommendation systems, and particularly relates to a knowledge graph recommendation method based on multi-view contrast learning. BACKGROUND

[0002] The core of a recommendation system is to recommend items that users are potentially interested in to users in a flood of information. As a classic recommendation algorithm, collaborative filtering has been widely studied and applied due to its effectiveness and universality. However, collaborative filtering cannot model edge information, which will lead to the problems of data sparsity and cold start. Researchers usually adopt the way of integrating edge information into the recommendation system to alleviate the problems of data sparsity and cold start in collaborative filtering, so as to improve the accuracy and interpretability of recommendation. Among various edge information, knowledge graph containing rich item information stands out. Knowledge graph is a kind of directed heterogeneous information network, which contains a large number of entities (for example: item entity, attribute entity, other entity) and relationships between entities. Knowledge graph is integrated into the recommendation system to obtain high-quality user and item representations. Existing knowledge graph-aware recommendation can be roughly divided into three categories: embedding-based recommendation, path-based recommendation, and propagation-based recommendation.

[0003] In the embedding-based method, knowledge graph embedding (KGE) method is applied to pre-train the knowledge graph to obtain item embedding in a specific semantic space. For example, CKE proposed by Zhang et al. uses TransR on the knowledge graph to obtain entity representation containing structural information, so as to obtain item representation aligned with it. DKN proposed by Wang et al. has two different channels of entity representation and word representation. First, DKN learns entity representation through TransD, and then key entities in news text are extracted. Finally, news representation is obtained by combining key entity representation and original word representation. KTUP proposed by Cao et al. is a translation-based model learned through TransH. Although the knowledge graph embedding method can obtain item representation containing structural information, compared with the recommendation task, the knowledge graph embedding method is more suitable for in-graph applications (for example: knowledge graph reasoning, knowledge graph completion), and the obtained item representation cannot represent the correlation between items.

[0004] In the path-based method, multiple connection patterns between users and items are mined based on the knowledge graph, providing accurate guidance for recommendation. The path-based method regards the knowledge graph as a directed heterogeneous information network, and extracts its potential features based on meta-path or meta-graph to represent the relevance between users and items. The path-based method can effectively improve the accuracy and interpretability of recommendation, but it needs to extract and construct a large number of high-quality meta-paths or meta-graphs from the knowledge graph to achieve the desired effect. In addition, the path-based method is not an end-to-end framework, which leads to the fact that we have to extract and construct meta-paths or meta-graphs based on new scenarios when the recommendation scenario changes. Constructing high-quality meta-paths or meta-graphs itself is a very complex task, which requires designers to have special rich domain knowledge.

[0005] The propagation-based method iteratively propagates on the knowledge graph to integrate domain information from different hops to enrich the representation of nodes. Its optimization objective is set to be consistent with the recommendation system to improve the effect of recommendation. RippleNet proposed by Wang et al. propagates the user's potential preferences on the knowledge graph, and uses the entity representation on the knowledge graph to update the user representation. However, RippleNet does not consider the importance of relationships in the propagation process, and the size of the ripple set constructed based on the knowledge graph may become out of control as the size of the knowledge graph increases, resulting in a large amount of storage and computing loss. KGCN proposed by Wang et al. only samples part of the domain information of entities to avoid introducing more noise, and then applies the framework of graph convolution network to propagate information to obtain item representations containing important domain information. However, KGCN does not consider explicit collaboration signals, which will cause information missing in the obtained item representation. KGAT proposed by Wang et al. is a model that propagates information on the collaborative knowledge graph to enrich the representation of users and items. However, KGAT does not distinguish between user nodes and item nodes in the information propagation process, which will result in more noise in the obtained user representation and item representation. KGIN proposed by Wang et al. models the user's intention behind the user-item interaction from a more fine-grained perspective, and executes GNN on the collaborative knowledge graph composed of the knowledge graph and the user-item interaction graph.

[0006] In general, the above proposed models are trained in a supervised learning manner, and the supervision signal only comes from sparse historical interactions. SUMMARY

[0007] To address the shortcomings of existing technologies, this invention provides a knowledge graph recommendation method based on multi-view comparative learning. First, it comprehensively considers and constructs four different views from different perspectives, including a collaborative knowledge graph jointly composed of a knowledge graph and a user-item interaction graph, a user-item interaction graph, and user-user and item-item graphs constructed based on the user-item interaction graph. Then, it performs comparative learning on the four views at both local and global levels, aiming to mine collaborative information between users and items, between users, and between items, as well as global structural information in a self-supervised manner, thereby alleviating the problem of sparse supervision signals.

[0008] To achieve the above objectives, this application proposes a knowledge graph recommendation method based on multi-view contrast learning, including:

[0009] Step 1: Build four different types of views

[0010] A collaborative knowledge graph containing global structural information is constructed by combining the knowledge graph and the user-item interaction graph; based on the user-item interaction graph, and by applying intersection-union ratio and receptive field, user-user graphs and item-item graphs are constructed in a reasonable manner.

[0011] Step 2: Comparative learning between local-level collaborative views

[0012] The architecture based on graph convolutional networks encodes three local collaborative views: the user-item interaction graph, the user-user graph, and the item-item graph.

[0013] Local-level contrastive learning is performed between user-item interaction graphs and user-user graphs, as well as between user-item interaction graphs and item-item graphs, to learn comprehensive user and item representations;

[0014] Step 3: Compare and learn between the global-level structure view and the local-level collaboration view.

[0015] The architecture based on graph neural networks encodes the global structural view of the collaborative knowledge graph (a combination of knowledge graph and user-item interaction graph);

[0016] Perform global-level comparative learning between global-level structure views and local-level collaboration views to learn discriminative user and project representations;

[0017] Step 4: Recommendation and Prediction

[0018] The final user and item representations are obtained by adding and concatenating user representations and item representations from different views;

[0019] The predicted matching score is obtained by using the final user embedding and item embedding representations, as follows:

[0020] inner product of the user embedding representation and the item embedding representation to predict a matching score thereof;

[0021] Step 5: optimization

[0022] The knowledge-aware recommendation task is optimized by a pair of BPR losses, specifically:

[0023] It assumes that the user's historical interactions indicate more user preferences, and should be assigned a higher prediction score than unobserved interactions;

[0024] Finally, the final target loss function is obtained by combining the local-level contrastive loss function, the global-level contrastive loss function and the knowledge-aware recommendation task loss function.

[0025] The method proposed in the application is tested on three data sets of Book-Crossing, MovieLens-1M and Last.FM, and in CTR prediction, the method has significant improvement compared with the existing most advanced baseline method, which is summarized as follows: (1) the method comprehensively considers four different views at the global and local levels, especially reasonably constructs the user-user graph and the item-item graph by applying the intersection-over-union and the receptive field; (2) the contrastive learning at the global and local levels of the method can fully capture the collaborative information between users and items, between users and between items, and the global structure information. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A multi-view contrastive learning-based knowledge graph recommendation method flowchart in the embodiment;

[0027] Figure 2 A user-user graph and an item-item graph example graph in the embodiment. DETAILED DESCRIPTION

[0028] The application will be further described below in combination with the drawings and embodiments;

[0029] The overall flowchart of the application is shown in Figure 1 .

[0030] The application comprises four components: a multi-view generation layer, a local-level contrastive learning layer, a global-level contrastive learning layer and a prediction layer.

[0031] The first, a multi-view generation layer, generates four different views from different perspectives, including a collaborative knowledge graph containing global structure information constructed by the knowledge graph and the user-item interaction graph, a user-item interaction graph containing user and item collaboration information, and a user-user graph containing user collaboration information and a project-project graph containing project collaboration information based on the user-item interaction graph. In this layer, the intersection over union is applied to reasonably construct the user-user graph and the project-project graph, and a receptive field is designed to avoid introducing more noise.

[0032] The second, a local-level contrastive learning layer, first encodes the user-item interaction graph, the user-user graph and the project-project graph based on the framework of the graph convolution network, and then performs contrastive learning between the user-item interaction graph and the user-user graph and between the user-item interaction graph and the project-project graph to learn comprehensive node representations at the local level.

[0033] The third, a global-level contrastive learning layer, first performs information propagation on the collaborative knowledge graph to obtain global user and project representations containing structure information, and then performs contrastive learning between the global-level structure view and the local-level collaboration view to learn discriminative node representations at the global level.

[0034] The fourth, a prediction layer, outputs the predicted interaction probability through the finally obtained user and project representations.

[0035] A knowledge graph recommendation method based on multi-view contrastive learning, specifically comprising the following steps:

[0036] Step 1: Construct four different types of views;

[0037] First, the knowledge graph G and the user-item interaction graph I are combined to construct a collaborative knowledge graph G containing global structure information c ; In addition, based on the user-item interaction graph, two local-level collaboration views are constructed, including a user-user graph containing user collaboration information and a project-project graph containing project collaboration information; as shown in Figure 2 Based on the user-item interaction graph, the set of items interacting with the user and the set of users interacting with the item can be obtained, and then the intersection over union is applied to obtain the relevance between users and between projects, as follows:

[0038]

[0039]

[0040] wherein, represents the relevance between users and users, represents the relevance between items, the numerator represents the size of the intersection of two sets, and the denominator represents the size of the union of two sets; u and u * represent different users, respectively, and v and v * represent different items; the intersection-union ratio can take into account the size of each set, so using the intersection-union ratio can reasonably and sufficiently obtain the relevance between two sets, and thus obtain the relevance between users or items behind two sets; as shown in Figure 2 , the user set interacting with item v1 is {u1, u2}, and the user set interacting with item v2 is {u1, u3}; the relevance between item v1 and item v2 can be calculated as , which can indirectly indicate that item v1 and item v2 have attribute similarity; the reason for not directly using the attribute set of the item in the knowledge graph to calculate is that the knowledge graph contains a large number of irrelevant attributes of the item, which will introduce a lot of noise, thereby greatly affecting the calculation result; similarly, the relevance between user u1 and user u3 can be calculated as , which can directly indicate that user u1 and user u3 have preference similarity;

[0041] Then a receptive field is designed to filter noise, thereby retaining the strongest user-user connection and item-item connection, as shown below:

[0042]

[0043]

[0044] wherein and are sparse graph adjacency matrices; top_K(IoU u ), top_K(IoU v ) respectively represent the K strongest connections associated with user u and item v; next, and are normalized as shown below:

[0045]

[0046]

[0047] wherein D is or diagonal matrix;

[0048] Step 2: Contrastive learning between local-level collaborative views;

[0049] The contrastive learning between local-level collaborative views includes two parts: encoding the local-level collaborative views and local-level contrastive learning optimization;

[0050] Step 2.1: Encoding the local-level collaborative views;

[0051] First, based on the architecture of the graph convolutional network, the user-item interaction graph is encoded, and the obtained user encoding and item encoding are represented as follows:

[0052]

[0053]

[0054] where N u and N v represent the neighbor set of user u and item v, respectively; after iterating B layers, considering the connectivity information of different orders, all order representations are added to obtain a single representation as follows:

[0055]

[0056]

[0057] where the size of B can be changed to control the strength of information propagation on the user-item interaction graph; then, recursive information propagation is performed in the user-user graph and the item-item graph, respectively, and the obtained user representation and item representation are as follows:

[0058]

[0059]

[0060] where N u represents the neighbor user set of user u in the user-user graph, N v represents the neighbor item set of item v in the item-item graph, M u and M v are the normalized sparse graph adjacency matrices obtained in step 1; after iterating L layers, all order representations are added to obtain a single representation as follows:

[0061]

[0062]

[0063] where the size of L can be changed to control the strength of information propagation on the user-user graph and the item-item graph;

[0064] Step 2.2: Local-level contrastive learning optimization;

[0065] between the specific representation of user u from the user-item interaction graph and the specific representation of user u from the user-user graph and between the specific representation of item v from the user-item interaction graph and the specific representation of item v from the item-item graph Contrastive learning is performed between them; first, a hidden layer in a neural network is applied to map them into the same contrastive space as follows:

[0066]

[0067]

[0068]

[0069]

[0070] where, respectively represent the representation of user u obtained by encoding the user-item interaction graph, the representation of item v obtained by encoding the user-item interaction graph, the representation of user u obtained by encoding the user-user graph, the representation of item v obtained by encoding the item-item graph; W1, W2 are trainable weight matrices; b1, b2 are biases; σ is a LeakyReLU activation function; the pair of representations of the same node obtained by pair view learning are taken as positive samples, and other forms of pair of representations are taken as negative samples, and the contrastive loss is defined as follows:

[0071]

[0072]

[0073] where and both represent positive representation pairs; and both represent negative representation pairs constructed in the same view; and both represent negative representation pairs constructed between different views; cos_sim represents the cosine similarity between two vectors; then the local level contrastive loss is obtained by jointly and as follows:

[0074]

[0075] Step 3: Contrastive learning between global level structure view and local level collaborative view;

[0076] Comparative learning between global-level structure view and local-level collaborative view contains two parts: encoding global-level structure view and global-level comparative learning optimization;

[0077] Step 3.1: Encoding global-level structure view;

[0078] A relation-aware graph neural network is designed to encode collaborative knowledge graph containing global structure information, and the resulting user and item representations are as follows:

[0079]

[0080]

[0081] Where π(v, r, t) represents the importance of neighbor entity t to the center entity v in the information propagation process based on the relationship r; π(v, r, t) is realized through a relation-based attention mechanism, which is expressed as follows:

[0082] π(v, r, t) = cos_sim(v + r, t)

[0083] Where the cosine similarity between v + r and t is used to model attention, which can fully preserve the structure information of the subgraph centered on v, so that neighbors closer to v get more attention; then the softmax function is used to normalize the coefficients of all neighbors adjacent to v:

[0084]

[0085] After iterating C layers on the collaborative knowledge graph, all-order representations are added to obtain a single representation, as follows:

[0086]

[0087]

[0088] Step 3.2: Global-level comparative learning optimization;

[0089] First, apply a hidden layer in a neural network to map global-level node representations and local-level node representations to the same contrastive space, as follows:

[0090]

[0091]

[0092]

[0093]

[0094] Then the same positive and negative sample strategy as local-level contrastive learning optimization is taken to define its contrastive loss as follows:

[0095]

[0096]

[0097]

[0098]

[0099] wherein, both represent positive representation pairs; both represent negative representation pairs constructed in the same view; both represent negative representation pairs constructed between different views; and and respectively represent the contrastive loss calculated from the global view and from the local view; then the joint and and and get the global-level contrastive loss:

[0100]

[0101] wherein M and N represent the number of positive representation pairs adopted in the corresponding loss function.

[0102] Step 4: recommendation prediction;

[0103] After encoding the four views, multiple representations of the user u and multiple representations of the item v can be obtained; then, by adding and connecting these representations to obtain the final representation of the user u and the item v as follows:

[0104]

[0105]

[0106] wherein || is the connection operator used for vector splicing; next, the inner product between the representation of the user u and the representation of the item v is performed to predict the matching score between the user u and the item v:

[0107]

[0108] Step 5: optimization;

[0109] For the knowledge-aware recommendation task, it is optimized by using a pair-wise BPR loss, which assumes that a user's historical interactions indicate more user preferences and should be assigned a higher prediction score than unobserved interactions:

[0110]

[0111] where O = {(u, v s , v p ) | (u, v s ) e O + , (u, v p ) e O -} denotes the training set, O + denotes the historical (positive) interactions of user u, O - denotes the unobserved (negative) interactions of user u; σ is a sigmoid function; and finally, the target loss function is obtained by combining L local , L global and L BPR :

[0112]

[0113] where Θ is the set of model parameters; α and β are hyperparameters that control the strength of different losses; is an L2 regularization term parameterized by λ to prevent overfitting of the model.

[0114] The method of the present application has carried out a large number of experiments based on three real-world data sets of Book-Crossing, MovieLens-1M and Last.FM. In CTR prediction, the trained model is used to predict each interaction in the test set, and two widely used evaluation indicators AUC and F1 are used to evaluate the CTR prediction. The CTR prediction results based on the three data sets are shown in Table 1. From the experimental results, it can be seen that the performance of the method of the present application is always better than the most advanced baseline on the three data sets. More specifically, the method of the present application achieves a relative improvement of 5.06% and 2.66% compared to the best baseline for AUC and F1 evaluation indicators on the Book-Crossing data set, respectively, and a relative improvement of 2.11% and 2.59% compared to the best baseline for AUC and F1 evaluation indicators on the MovieLens-1M data set, respectively, and a relative improvement of 4.08% and 6.45% compared to the best baseline for AUC and F1 evaluation indicators on the Last.FM data set, respectively.

[0115] Table 1 AUC and F1 results under CTR prediction based on three data sets

[0116]

Claims

1. A knowledge graph recommendation method based on multi-view contrastive learning, characterized in that, Specifically comprising the following steps: Step 1: Constructing four different types of views Joint knowledge graph and user-item interaction graph construct collaborative knowledge graph containing global structure information; Based on user-item interaction graph, and apply intersection over union and receptive field to construct user-user graph and item-item graph; Step 2: Contrastive learning between local level collaborative views Based on the architecture of graph convolutional network, encode user-item interaction graph, user-user graph and item-item graph, three local collaborative views; Local level contrastive learning between user-item interaction graph and user-user graph, and between user-item interaction graph and item-item graph, to learn comprehensive user and item representation; Step 3: Contrastive learning between global level structure view and local level collaborative view Based on the architecture of graph neural network, encode collaborative knowledge graph, a global level structure view; Global level contrastive learning between global level structure view and local level collaborative view, to learn discriminative user and item representation; Step 4: Recommendation prediction By adding and connecting operation between user representations and item representations from different views to get the final user and item representation; By the final user embedding representation and item embedding representation, get the predicted matching score, specifically: inner product of user embedding representation and item embedding representation to predict the matching score; Step 5: Optimization Optimize knowledge-aware recommendation task through pairwise BPR loss, specifically: it assumes that user's historical interaction indicates more user preference, which should be assigned a higher predicted score than unobserved interaction; Finally, get the final target loss function through joint local level contrastive loss function, global level contrastive loss function and knowledge-aware recommendation task loss function.

2. The method of claim 1, wherein the method is based on multi-view contrastive learning. The step 1 is specifically as follows: First joint knowledge graph and user-item interaction graph Constructing a collaborative knowledge graph containing global structure information Constructing two local-level collaborative views, a user-user graph containing collaboration information between users and a project-project graph containing collaboration information between projects, based on the user-item interaction graph Based on user-item interaction graph, get the set of items interacted with users and the set of users interacted with items, and then apply intersection over union to get the relevance between users and between items, as follows: ； ； wherein, denotes the association between users, denotes the association between items, the numerator denotes the size of the intersection of the two sets, and the denominator denotes the size of the union of the two sets; and denote different users, and denote different items; Then a receptive field is designed to filter noise, so as to retain the strongest user-user connection and item-item connection, as follows: ； ； wherein and are sparse graph adjacency matrices; , denote the top and most relevant connections to users and items , are normalized as follows: ； ； wherein is or a diagonal matrix.

3. The method of claim 1, wherein the method is based on multi-view contrastive learning. The step 2 is specifically as follows: Contrastive learning between local level collaborative views includes two parts: encoding local level collaborative views and local level contrastive learning optimization; Step 2.1: Encoding local level collaborative views; First, based on the architecture of graph convolutional network, encode user-item interaction graph, and the obtained user encoding and item encoding are as follows: ； ； where and denote the neighborhood sets of users and items respectively; after iterating layers, considering the connectivity information of different orders, all orders' representations are added to get a single representation as follows: ； ； where the size of can control the strength of information propagation on the user-item interaction graph; then recursively perform information propagation in the user-user graph and item-item graph respectively, and the resulting user representation and item representation are shown as follows: ； ； where representing users a set of neighbor users in the user-user graph, representing items a set of neighbor items in the item-item graph, and are the normalized sparse graph adjacency matrices obtained in step 1; after iteration layers, all orders of representation are added to obtain a single representation as follows: ； ； where the size of the can control the strength of information propagation on the user-user graph and the item-item graph; Step 2.2: Local level contrastive learning optimization; In the user-project interaction diagram Specific representation And users from the user-user graph Specific representation Between, and in the project from the user-project interaction diagram Specific representation And projects from the project-project diagram Specific representation They are compared and learned from each other; firstly, a hidden layer in a neural network is applied to map them to the same contrast space, as shown below: ； ； ； ； where, , , , respectively represent the representation of users obtained by encoding the user-item interaction graph, the representation of items obtained by encoding the user-item interaction graph, the representation of users obtained by encoding the user-user graph, the representation of items obtained by encoding the item-item graph; , is a trainable weight matrix; , is a bias; is a LeakyReLU activation function; the pair-wise representation of the same node obtained by pair-wise view learning is regarded as a positive sample, and other forms of pair-wise representation are regarded as negative samples, and the contrastive loss is defined as follows: ； ； where and both represent positive pairs of representations; and both represent negative pairs of representations constructed in the same view; and both represent negative pairs of representations constructed between different views; represents the cosine similarity between two vectors; then the local-level contrastive loss is obtained by jointly and as follows: 。 4. The method of claim 1, wherein the method is based on multi-view contrastive learning. The step 3 is specifically as follows: Contrastive learning between global level structure view and local level collaborative view includes two parts: encoding global level structure view and global level contrastive learning optimization; Step 3.1: Encoding global level structure view; A relation-aware graph neural network is designed to encode collaborative knowledge graph containing global structure information, and the obtained user and item representation are as follows: ； ； wherein represents a relationship-based attention mechanism the importance of the neighbor entities for the center entity is achieved by the relationship-based attention mechanism which is expressed as follows: ； where the cosine similarity between and is used to model the attention, can preserve the structure information of subgraphs centered at well, so that neighbors closer to get more attention. Then the softmax function is used to normalize the coefficients of all neighbors of the current pixel All neighboring coefficients of the adjacent neighbors are normalized: ； Iterating on a collaborative knowledge graph After the layers, all the stages' representations are added to obtain a single representation, as follows: ； ； Step 3.2: Global level contrastive learning optimization; First, we apply a hidden layer in a neural network to map the global-level node representations and the local-level node representations into the same contrastive space, as follows: ； ； ； ； Then we adopt the same positive-negative sample strategy as the local-level contrastive learning optimization to define its contrastive loss, as follows: ； ； ； ； wherein, , , , all represent positive representation pairs; , , , all represent negative representation pairs constructed in the same view; , , , all represent negative representation pairs constructed between different views; , respectively represent the contrastive loss computed from the global view and the local view; then the global level contrastive loss is obtained by jointing and , ​ ； wherein and denotes the number of positive representation pairs employed in the respective loss function.

5. The method of claim 1, wherein the method is based on multi-view contrastive learning. The step 4 is specifically as follows: After encoding the four views, the user is obtained. Multiple representations and projects Multiple representations Then, the final user representation is obtained by adding and concatenating these representations. and projects The representation is as follows: ； ； in This is the concatenation operator used for vector concatenation; next, we will perform user... The representation of and projects The representation of The inner product between them is used to predict users. and projects Match score between: 。 6. The method of claim 1, wherein the method is based on multi-view contrastive learning. The step 5 is specifically as follows: For the knowledge-aware recommendation task, we optimize it with a pairwise BPR loss, specifically, the user's historical interactions indicate more user preferences, which should be assigned a higher prediction score than the unobserved interactions: ； where denotes the training set, denotes the user 's historical positive interactions, denotes the user 's unobserved negative interactions; is the sigmoid function; finally, the target loss function is obtained by jointly , and ​ ； in It is the model parameter set; and It is a hyperparameter that controls the intensity of different losses; This is set to prevent model overfitting. A parameterized one Regularization term.

Images