Recommendation Model Research Based On Heterogeneous Graph

The collaborative filtering recommendation algorithm only mines the user’s preference information from the historical interaction data between the user and the item,which has the advantage of simple calculation,but has the problem of data sparsity,that is,there are a large number of zero or missing values in the interaction data between users and items.The current research trend is to introduce auxiliary information into the recommendation algorithm to solve the above problem,thereby improving the recommendation performance.Knowledge graph is a heterogeneous information network graph,which contains a large amount of relational semantic information,which can make up for the sparseness or lack of user-item interaction data,so it can be integrated into the recommendation system as auxiliary information.In this paper,a research on the recommendation model based on heterogeneous graph were carried out,and its main work and innovations are as follows:Aiming at the data sparsity problem of traditional collaborative filtering recommendation algorithm,a neighbor information aggregation recommendation model based on knowledge graph were proposed in this paper.On the one hand,the model matches the user’s historical preferences and candidates item with entities in the knowledge graph,and acts as a central node to expand outward to more neighbor entities along different link relationships,and aggregates neighbor information with graph convolutional network to enrich the feature representation of user and item;on the other hand,considering that if all neighbor information is aggregated indiscriminately,the feature information of the central node will easily lack representation ability.We introduce self-attention mechanism to focus on the link relationship between nodes,so that neighbor information of nodes is aggregated biasedly.The model is experimentally verified on two public datasets in this paper.The results show that the proposed model has a great improvement in the recommendation accuracy compared with the existing recommendation models in the CTR prediction and Top-K recommendation tasks.Aiming at the problem of data sparsity in recommender system,the recommendation model based on recurrent neural network and knowledge graph were proposed in this paper.Considering that existing graph embedding-based recommendation models ignore the importance of user-item interactions,the learned embedding representation cannot effectively express user interests.Firstly,the user’s local interest features and item latent features are extracted from the user’s historical interaction sequence through an improved recurrent neural network,which is beneficial to obtain real user preferences.Secondly,in order to transfer the entity information in knowledge graph to the original features of users and items,a new entity-aware embedding method is proposed,which directly integrates the collaborative information between users and historical items into entities to explores the user’s interest features from a global perspective.Finally,the recommended tasks is completed by integrating the user’s local interest features and global interest features.We conducted experiments on two real data sets and the experimental results prove that the proposed model has a higher recommendation accuracy compared with the existing recommendation models.