Item-Behavior Sequence Session-Based Recommendation

Abstract

AbstractSession-based Recommendation has become important and popular in Recommendation System, which focuses on leveraging the historic records to predict the next interactive item(s). The previous works only employ the item sequence or process item sequence and users’ behavior respectively. To model session with item sequence and behavior sequence simultaneously, we propose the Item-Behavior Sequence Session-based Recommendation (IBSSR for abbreviation). The Hadamard product of item sequence and behavior sequence is fed into an improved Transformer and behavior sequence is fed into a Gated Recurrent Unit. Representation generated by concatenating these embeddings with soft attention mechanism will be used to predicts the next item. To verify the bound of our model, we conducted experiments on two datasets where there are large differences in ability of behavior contacting context. Experiment proofs that our model can effectively model item sequence whose behavior sequence has strong ability in contacting context based on the session.KeywordsSequential recommendationFine-grained behaviorResidualTransformerGRU