SimGNN: simplified graph neural networks for session-based recommendation

Session-based recommender systems (SBR) aim to predict the next action of an anonymous user session. Recently Graph Neural Networks (GNN) models have gained a lot of attention in this task. Existing models learn sequential complex transition patterns using the Gated Graph Neural Networks (GGNN) architecture. We argue that learning non-sequential complex transition patterns may be sufficient in SBR due to the short time interval and length of the sessions. To fully exploit the advantages of non-sequential GNN such as scalability, we design Simplified Graph Neural Network for Session-based Recommendation SimGNN, a non-sequential, linear GNN model for interaction representation. SimGNN uses the k-th power of the normalized adjacency matrix and the current session interactions to learn the k-th layer interaction representation. To improve the representation, SimGNN uses a highway gating mechanism. From the interaction representation learned by the proposed non-sequential and linear model, SimGNN models local preference and global preference and uses a proposed gating mechanism to aggregate these preferences. Experimental results showed that SimGNN outperforms state-of-the-art sequential GGNN models for SBR in terms of accuracy metrics - precision and mean reciprocal ranking.