Quaternion-based knowledge graph neural network for social recommendation

Abstract

In recent years, the surge in the number of users in recommender systems has brought unprecedented opportunities and challenges to research on recommender systems. The development of the graph neural network makes social recommendation based on graph data more advantageous, but it also brings problems with parameterization. Although the recommender system contains a large amount of data, such as users, items, and user-item interactions, the social relationship data of users in social recommendation are still sparse. To solve the problem of parameterization, we propose a novel quaternion-based knowledge graph neural network for social recommendation (QSoR), which applies hypercomplex space to the social recommendation. Without affecting the recommendation quality, QSoR utilizes the expressibility of the quaternion and weight sharing mechanism of the Hamilton product to reduce the parameters during training. To solve the problem of the sparsity of users’ social relationship data, we propose a quaternion-based explicit and implicit social relationships integration algorithm, which obtains users’ implicit social relationships from the user-item graph as a supplement for users’ social information. We design a double-end attention mechanism when obtaining the user’s preferences and social relationships, which makes full use of the item information and the interaction information between the user and the neighbor nodes. In this way, we can obtain users’ personalized preferences, distinguish the importance of user relationships, and obtain neighbors with a high degree of influence on users, thus improving the accuracy of recommendation results. The social relationship graph and the user-item graph are effectively combined in the quaternion space to make personalized recommendations for users. Experiments on three public datasets, Yelp, Last-FM, and Epinions, demonstrate the validity of the proposed model QSoR.