MGU-GNN: Minimal Gated Unit based Graph Neural Network for Session-based Recommendation

Session-based recommender systems (SBRS) play a crucial role in predicting the next click of a user from anonymous session data on numerous online platforms such as e-commerce, music, etc. However, predicting the next click is a very challenging task within the session, as it contains a very little amount of contextual information. Most of the existing techniques consider a session like a sequence of items to make recommendations and ignore the complex transition between items. In order to get accurate item embeddings and capture the complex transitions of items, we have proposed a Minimal Gated Unit based Graph Neural Network (MGU-GNN) for the session-based recommendation (SBR) tasks. We have also integrated a soft-attention network and target-based interest-aware network module, called MGU-GNN-TAR. The target-based interest-aware network module adapts to varying users' interests in terms of the items to be targeted. The soft-attention network module adapts long-term priorities and current session interest for better prediction of the user's next item or action. This model provides precise item embedding by incorporating the complex item transitions. The proposed model uses a gated mechanism called the Minimal Gated Unit, which has a single gate, and due to this reason, the parameters have been reduced to 67% as compared to the GRU cell. A GRU cell is the most basic of all gated hidden units. To demonstrate the efficacy of the proposed models, comprehensive experiments on four most commonly used publicly available real-world datasets have been performed, and they show that the proposed models routinely beat baseline methods and state-of-the-art SBR techniques on all four datasets.