Collaborative filtering: Graph neural network with attention

Abstract

In recent years, graph neural networks have been widely used in natural language processing and speech recognition. However, there is relatively little exploration of graph neural networks in recommendation systems. In this work, we strive to develop neural network based technology to solve the problem of collaborative filtering recommendation based on implicit feedback. Some recent work use deep learning for recommendation, but they mainly use it for auxiliary information modeling. When modeling the key elements of collaborative filtering (the interaction between users and item features), they still use matrix factorization and apply inner products to the potential features of users and items. The collaboration signal(interaction information between users and items) will not be encoded during the embedding process. At the same time, the high-order connectivity of user-item cannot be fully utilized, making the interactive information not comprehensive enough. Therefore, the final embedding may not be enough to capture the collaborative filtering effect. By replacing the inner product with a neural architecture that can learn arbitrary functions from data, we propose a general method called “Graph Neural Network with Attention” (GNNA). GNNA captures CF signals based on neural networks and uses high-order connectivity to obtain neglected interactive information, thereby improving the embedding of users and items. By introducing attention mechanism and high-order connectivity, it can learn user vectors and item vectors on the user item interaction graph, collect neighbor interaction information for coding, and spread it on the user-item interaction graph. This can comprehensively inject user-item collaboration signals into the embedding process. We have conducted extensive experiments on two public benchmarks. Further analysis verified the importance of using attention mechanism and high-order connectivity to learn user and item representations, and proved the rationality and effectiveness of GNNA.