Multi-behavior recommendation method based on knowledge graph time projection

Abstract

The invention discloses a multi-behavior recommendation method based on knowledge graph time projection, and the recommendation method introduces a knowledge graph of time factors to accurately capture dynamic changes of user interests and behaviors and timeliness characteristics of commodities, thereby improving the perception ability of a recommendation model in the time dimension. Specifically, the model integrates time-aware knowledge graph representation on the basis of multi-behavior representation to improve the recommendation performance of the multi-behavior recommendation method. Performing preliminary modeling on representation of the user and the project by taking the recommendation task as a main supervision task; time projection is carried out by using information of a knowledge graph to learn time-perceived correlation embedded representation between projects, so that fine-grained intention differences between different behaviors of a user and timeliness characteristics between articles are fused into multi-behavior embedded representation of each interaction relationship; and the multi-behavior embedding of the project and the user can be more efficiently and accurately modeled. After the auxiliary behaviors and the target behaviors of the user and the project are accurately modeled, supervision signals of the auxiliary behaviors are transferred to the target behaviors through comparative learning, so that the recommendation performance of the model to the target behaviors of the user is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of graph neural network, recommendation system, knowledge graph, contrastive representation learning, etc., and specifically designs a multi-behavior recommendation method based on knowledge graph to generate time projection embedding

[0002] 3. Background technology of the invention

[0003] With the popularization and development of the Internet and social media, the speed of information generation and dissemination has far surpassed any previous era. Although the rapid development of the Internet has greatly facilitated people's daily lives, it has also inevitably caused the problem of information overload, making users often feel at a loss when faced with a large amount of content. In this case, it is often difficult for users to quickly filter out valuable information in a limited time, and merchants are also faced with the challenge of how to effectively attract users' attention and accurately grasp their needs in fierce competition. Therefore, as a key technology to deal with information overload, the recommendation system can deeply explore users' interests and preferences by analyzing users' historical data, interaction records and behavior trajectories, so as to predict the content or products that they may be interested in in the future. For this reason, the recommendation system is now widely used in e-commerce platforms, online video websites, social media and other fields to help users quickly locate content of interest, which not only effectively alleviates the problem of information overload, but also improves user experience and enhances the market competitiveness of the platform.

[0004] Among many recommendation technologies, graph convolutional networks are widely used in single-behavior based user-item interaction recommendations because they can inject high-order connection signals into collaborative filtering and integrate multi-hop neighbors into node representation learning. Traditional collaborative filtering technology usually assumes that users have only a single type of behavior on items. However, in real recommendation scenarios, user interactions include not only purchase behaviors, but also multiple ways of interacting with items such as page browsing, adding to shopping carts, and liking. These behaviors can provide rich insights for user preference learning. Therefore, the multi-behavior recommendation system proposes to treat these interactive behaviors other than purchase behaviors as auxiliary behaviors to provide information enhancement for the model, helping the model to better model and predict the user's target behavior (that is, the user's purchase behavior).

[0005] In order to further utilize the behavioral relationships between users and items contained in multi-behavior information. Thanks to the fact that knowledge graphs can well construct and represent the inherent facts of interactions between users and items and the potential semantic relevance between items. Therefore, many studies have attempted to use knowledge graphs to capture the interdependencies between different behaviors and distinguish the personalized preferences of different users, so as to model user preferences from multiple perspectives and enhance target behavior recommendations.

[0006] 4. Summary of the invention (detailed description of the technical solution of the present invention, if there are drawings, should be combined with flow charts, block diagrams, circuit diagrams, timing diagrams, structure diagrams, etc.)

[0007] 1. Technical Problems Solved

[0008] In real-world scenarios, user interests and behaviors are time-sensitive, and behaviors in different time periods may show significant differences. For example, the popularity of certain products may change with the seasons, and users' interests may shift in different periods. However, mainstream knowledge graph embedding methods ignore the availability and importance of time range when learning the embedding of entities (including users and items) and interactive behavior relationships in knowledge graphs. These methods treat knowledge graphs as static graphs and assume that the user intentions and behaviors contained therein are not time-sensitive, which ignores the timeliness of user intentions and behaviors, causing the model to learn one-sided representations and severe performance degradation.

[0009] In view of the shortcomings of the traditional multi-behavior recommendation model mentioned above, this paper proposes a graph neural network model. Based on the existing project relationship knowledge graph, a time factor is introduced to measure the strength of the association between users and projects. By projecting the embedded representation of users, items, and interaction relationships onto the time plane of the corresponding user historical interaction projects, the strength of the association between entity embeddings is dynamically weighed. Finally, it is applied to multi-behavior learning to improve the model's ability to represent user behavior intentions and recommendation accuracy.

[0010] 2. Technical Solution

[0011] Step 1: Obtain a public multi-behavior recommendation system dataset that covers multiple types of user behaviors, such as product purchases, adding to shopping carts, favorites, and browsing. Preprocess the downloaded data to filter out users with too low a frequency of interaction with the system, and renumber the remaining users and items. In terms of data set division, a time-aware leave-one-out evaluation method is used: the last interaction record of each user is divided into a test set, and the remaining interaction records are used as training sets. After processing, output the training set and test set files.

[0012] Step 2: Generate a relational knowledge graph between items based on the given training set, which includes the head entity, relationship, tail entity, and the average interval time between user interactions between the head entity and the tail entity.

[0013] Step 3: Import the generated project relationship knowledge graph data, project the embeddings of entities and relationships into the corresponding time embeddings, and generate time-specific entity and relationship embeddings. Use the distance translation model and semantic matching model to optimize the embedding representations of entities and relationships, and obtain the final time-projected embedding generated based on the knowledge graph through back-propagation learning (where the relationship between the head and tail entities is specific to the interaction relationship and time).

[0014] Step 4: After completing the entire process of step 3, import the preprocessed multi-behavior dataset. Use graph convolution technology to perform high-order neighborhood aggregation and learn the user's historical interaction data under each behavior. Generate embedded representations of users and items from multiple behavior perspectives by capturing the high-order connectivity between users and items.

[0015] Step 5: Use data augmentation technology to integrate the time projection embedding representation generated based on the knowledge graph into the multi-behavior embedding representation, so as to integrate the fine-grained intention differences between users in different behaviors and the timeliness characteristics between items into the multi-behavior embedding representation of each interaction relationship. And use the mutual information lower bound to transfer the semantics of the user's auxiliary behavior to the target behavior, so as to enrich the user semantic information of the target behavior and alleviate the sparse supervision signal problem of the target behavior.

[0016] Step 6: Calculate the loss value of the model based on the tasks obtained from steps 4 to 5, assign appropriate weights to each task loss for back propagation, and calculate the gradient value.

[0017] Step 7: Based on the parameter information of the gradient optimization model, the model is prompted to improve the novelty of the recommendation without losing the accuracy of the recommendation. By calculating the results of the model under different normalization coefficients and selecting the optimal model parameters based on the evaluation indicators, the Top-K project recommendations are finally generated for the user.

[0018] 3. Beneficial Effects

[0019] Compared with the existing technology, the present invention has the following advantages:

[0020] 1. It takes into account the user's multiple behavioral interaction information. Compared with the single-target behavior recommendation model, the multi-behavior recommendation model can significantly improve the recommendation performance of the user's target behavior, thanks to the fact that the auxiliary behavior has richer user behavior intention supervision signals.

[0021] 2. Introducing a knowledge graph that includes time factors and generating an embedded representation of the knowledge graph through time projection can not only more accurately capture the dynamic changes of user interests and behaviors, but also effectively characterize the timeliness characteristics of products, thereby improving the model's perception ability in the time dimension.

[0022] 3. By integrating the time projection embedding representation generated based on the knowledge graph into the multi-behavior embedding representation, the model can incorporate this part of the time dimension features into the multi-behavior representation, thereby more accurately modeling the user's behavioral intentions.

[0023] 4. By transferring the semantic information from the auxiliary behavior to the target behavior through the mutual information lower bound, the data sparsity problem under the target behavior can be alleviated, thereby improving the recommendation performance of the model.

[0024] 5. Specific implementation methods of the invention (proving that the technical solution can achieve the aforementioned beneficial effects through different specific experimental methods and the experimental data obtained, or explaining through principles that the aforementioned technical solution can solve the aforementioned technical problems and achieve the aforementioned technical effects) BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Overall flow chart of the recommended method

[0026] Figure 2 Model framework diagram of the present invention

[0027] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in combination with specific examples and with reference to the accompanying drawings.

[0028] The present invention recommends items based on user multi-behavior interaction data. The overall process is as follows: Figure 1 The specific steps include:

[0029] Step 1. Download the Tmall, IJCAI and Retail multi-behavior recommendation datasets publicly available on the Internet. The datasets mainly contain the historical interaction information of the four behaviors of users: browsing, adding to shopping carts, collecting and purchasing. By preprocessing the multi-behavior dataset, users with too few interactions are filtered out, and the users and items in the dataset are renumbered. Finally, the time-based leave-one-out evaluation method is used to generate the corresponding training set and test set files, that is, the test set only contains the user's last interaction item, and the rest of the interaction data is used for training. Each processed dataset will generate 5 txt files: one containing the user's historical interaction information files under four different behaviors (click.txt, cart.txt, favorite.txt, buy.txt) and a test set file (test.txt). In the historical interaction information files under these four different behaviors, each line consists of the user ID and the item ID that the user has interacted with under the corresponding behavior. In the test set file, the first value of each line represents the user's ID, and the second value is the item ID purchased by the user for the last time.

[0030] Step 2: Generate a relationship knowledge graph file (kg.txt) between items based on the given training set, in which the item ID of the target behavior is used to generate the head entity, the four behaviors are used as interaction relationships, the item ID of the auxiliary behavior is used as the tail entity, and the difference between the average time of all users interacting with the head entity under the target behavior and the average time of interacting with the tail entity under the auxiliary behavior is used as the time value of the tuple.

[0031] Step 3: Import the generated knowledge graph file (kg.txt) of the relationship between items. We regard time as a hyperplane, that is, for a knowledge graph with D time intervals, there are D different time hyperplanes, represented by the normal vector Therefore, based on the hyperplane, the knowledge graph of the relationship between items can be decomposed into several static subgraphs: We can combine the three groups G τ = {(h, r, t)} projected onto a specific time hyperplane w τ On. Triple embedding representation (e h ,e r ,e t ) can be expressed as:

[0032]

[0033] where ||w t ||2=1, we learn the embedding representation of entities and relations together For each time interval τ in .

[0034] Step 4. In order to accurately capture the correlation between items, we propose two time-projected optimization functions, TP-TransR and TP-TATEC, to model the correlation between items from multiple perspectives. TP-TransR can dynamically learn entity representations in the relational space, allowing the head (tail) entity to reflect the relevant relationship characteristics of its corresponding tail (head) entity in the knowledge graph under different relations and times. Similarly, TP-TATEC can capture the potential semantics of entities changing over time, thereby modeling the association relationship between entities. Therefore, for the quadruple (h, r, t, τ), the two optimization functions TP-TransR and TP-TATEC are defined as follows:

[0035]

[0036] in is the projection matrix of relation r, which projects entities from entity space to relation space. And D is a diagonal matrix shared by all interaction relations.

[0037] Step 5: Based on the time projection embedding and optimization functions of Steps 3 and 4, the model uses the constructed knowledge graph to learn the time-aware related embedding representations between projects from the perspectives of TP-TransR and TP-TATEC through back propagation (P τ (e h ),P τ (e r ),P τ (e t )).

[0038] Step 6: Import the preprocessed multi-behavior data, use the three-layer graph convolution architecture to perform neighborhood aggregation, and map the specific behavior interaction graph. The high-order neighbors of are injected into the embedding representation of users and items to enhance the information transfer between nodes. The process of neighborhood aggregation can be expressed as:

[0039]

[0040] in and They represent the refined embeddings of user u and item i under behavior k after l layers of propagation. represents the set of interaction items observed by user u in behavior k, Table 1 shows the set of users that have interacted with item i in behavior k. After l layers of graph convolution, we use mean pooling to further combine the embeddings obtained in each layer to form the final representation of a user (item):

[0041]

[0042] Step 7. Since multi-behavior embedding relies on the knowledge graph embedding learned in step 5 to further learn the time-aware correlation between items. Therefore, we use mutual information to quantify the similarity between multi-behavior embedding and knowledge graph embedding, and integrate the time projection embedding representation generated based on the knowledge graph into the multi-behavior representation, so as to integrate the fine-grained intention differences between users in different behaviors and the temporal characteristics between items into the multi-behavior embedding representation of each interaction relationship. Specifically, the head entity of the knowledge graph originates from the target behavior item, so we use the head entity of the knowledge graph to learn the temporal characteristics between items under the target behavior. The formula is as follows:

[0043]

[0044] Among them, t is a hyperparameter, which is represented as a temperature coefficient in softmax. And s(·) is the cosine similarity function. Secondly, given that the tail entity is defined by a specific relationship in the knowledge graph, we use the GCN aggregator to aggregate the tail entity representation e tand its associated relationship r As a representation of its specific relationship r,t :

[0045] e r,t =ReLU(W(e r +e t )b,

[0046] Among them, we set the activation function to the ReLU function, and W is a trainable weight matrix used to extract information useful for graph propagation. Similarly, we apply the tail entity set to the auxiliary behavior a to learn the inter-item temporal characteristics of the auxiliary behavior:

[0047]

[0048] Therefore, the final objective function for learning the time-aware correlation between knowledge graph embeddings is expressed as follows:

[0049] After learning the time-aware correlation between knowledge graph embeddings, the temporal semantic information in the auxiliary behavior and the target behavior is fully preserved. Therefore, in order to enrich the supervision signal in the target behavior, we perform contrastive learning to transfer the supervision signal of the auxiliary behavior to the target behavior to alleviate the data sparsity problem of the target behavior. We choose to regard the same node embeddings from different behaviors as positive pairs and different node embeddings from different behaviors as negative pairs. Taking the project side as an example, the project side optimization objective function based on InfoNCE can be expressed as:

[0050]

[0051] Combined with the user-side loss of similar design, the final objective optimization function of contrastive learning is:

[0052] Step 8. We use Bayesian Personalized Ranking (BPR) loss as the loss of the main supervision task to encourage the probability estimate of the user's observed interaction to be higher than the probability estimate of its unobserved interaction. The BPR loss under behavior k can be calculated as follows:

[0053]

[0054] Among them, σ is the Sigmoid nonlinear activation function. is the inner product of the final representation of the user and item under k behaviors. Θ represents a learnable parameter. λ controls the regularization strength. The overall BPR loss is the sum of the BPR losses of each behavior: We adopt a multi-task training strategy to jointly optimize the main supervision task loss, the time-aware correlation learning loss, and the inter-action contrastive learning loss:

[0055] Step 9: We use the optimizer to optimize the model parameters according to the gradient calculated in step 8 to ensure the accuracy of the model recommendation. We select the optimal model parameters based on the two evaluation indicators HR (Hit Ratio) and NDCG (Normalized Discounted Cumulative Gain), and finally generate Top-K project recommendations for users.

[0056] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A multi-behavior recommendation method based on knowledge graph time projection, whose features include Follow these steps: Step 1: Download the public multi-behavior recommendation dataset on the Internet. The dataset mainly contains the historical interaction information of the four behaviors of users: browsing, adding to shopping cart, collecting and purchasing. By preprocessing the multi-behavior dataset, users with too few interactions are filtered out, and the users and items in the dataset are renumbered. Finally, the time-based leave-one-out evaluation method is used to generate the corresponding training set and test set files, that is, the test set only contains the user's last interaction item, and the rest of the interaction data is used for training. Step 2: Generate a relationship knowledge graph file between items based on the given training set, in which the item ID of the target behavior is used to generate the head entity, multiple interactive behaviors are used as the relationship of the knowledge graph, the item ID of the auxiliary behavior is used as the tail entity, and the difference between the average time of all users interacting with the head entity under the target behavior and the average time of interacting with the tail entity under the auxiliary behavior is used as the time value of the tuple. Step 3: Import the generated knowledge graph file of the relationship between items. Treat time as a hyperplane, and project the head entity, relationship, and tail entity of the knowledge graph onto a specific time hyperplane. Then, the model uses the constructed knowledge graph to learn the time-aware embedding representation between items from the perspectives of TP-TransR and TP-TATEC through back propagation. Step 4: Import the preprocessed multi-behavior data, use the three-layer graph convolution architecture to perform neighborhood aggregation, and inject the high-order neighbors of the specific behavior interaction graph into the embedding representation of users and items to enhance the information transfer between nodes. And use mean pooling to further combine the embeddings obtained in each layer to obtain the final multi-behavior embedding representation. Step 5: Incorporate the time projection embedding representation generated based on the knowledge graph into the multi-behavior representation, so as to integrate the fine-grained intention differences between users in different behaviors and the timeliness characteristics between items into the multi-behavior embedding representation of each interaction relationship. And further use contrastive learning to transfer the supervision signal of the auxiliary behavior to the target behavior to alleviate the data sparsity of the target behavior. Step 6: Calculate the loss function of the model based on the tasks obtained in step 5, and adopt a multi-task training strategy to jointly optimize the main supervision task loss, time-aware correlation learning loss, and inter-behavior contrast learning loss, and perform back-propagation gradient calculation. Step 7: Optimize the model parameters based on the gradient to improve the accuracy of the model recommendation. Select the optimal model parameters based on the evaluation indicators and finally generate Top-K project recommendations for users.

2. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 3, the triplet embedding representation (e h ,e r ,e t ) is expressed as follows: where ‖w τ ‖2=1.

3. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 3, the two optimization functions TP-TransR and TP-TATEC are defined as follows: in is the projection matrix of relation r, which projects entities from entity space to relation space. And D is a diagonal matrix shared by all interaction relations.

4. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 4, the three graph convolutions performed on the interaction graph between user u and item i are expressed as follows: in They represent the refined embeddings of user u and item i under behavior k after l layers of propagation. represents the set of interaction items observed by user u in behavior k, Table 1 is the set of users that interacted with item i in behavior k.

5. According to claim 1, the multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 4, the multi-behavior embedding representation of the user (item) is defined as follows:

6. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 5, the formula for learning the temporal characteristics between items under the target behavior through the head entity of the knowledge graph is expressed as follows: Among them, t is a hyperparameter, which is represented as a temperature coefficient in softmax, and s(·) is the cosine similarity function.

7. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 5, the tail entity t under a specific relation r is represented as follows: have been r,t =ReLU(W(e r +e t )), Among them, we set the activation function to the ReLU function, and W is a trainable weight matrix used to extract information useful for graph propagation.

8. According to claim 1, a multi-behavior recommendation method based on knowledge graph time projection is characterized in that: In step 5, the formula for learning the inter-item temporal characteristics under auxiliary behavior a through the tail entity of the knowledge graph is expressed as follows:

9. The multi-behavior recommendation method based on knowledge graph time projection according to claim 1, characterized in that: In step 5, the project-side optimization objective function based on InfoNCE is defined as follows:

10. The multi-behavior recommendation method based on knowledge graph time projection according to claim 1, characterized in that: In step 6, the loss function of the multi-task training strategy is expressed as:

Images