Context enhanced multi-level attention model for session-based recommendation
Session-based recommendation aims to predict users'next click based on their interactions in anonymous sessions.While graph neural network(GNN)-based methods have shown promising results,they still have limitations.GNN-based methods overlook the session's sequential patterns and only consider transition patterns between items.Furthermore,most methods focus solely on the internal information of the current session and ignore external collaborative information from neighboring sessions,i.e.,contextual patterns.To address these issues,we propose a context enhanced multi-level attention model(CEMA),which uses multi-level attention mechanisms to learn item features and model user preferences at both item and session levels.CEMA applies a multi-layer GraphSAGE to learn complex transition patterns between items to capture local user preferences.The item-level attention mechanism in CEMA employs a gate attention unit to calculate item importance,identify user interests,and avoid noise interference.This helps capture the sequential patterns in the session to model users'global preferences.Moreover,a session-level attention mechanism is designed to efficiently calculate the similarity between sessions and extract contextual patterns to predict users'next click.Experiments conducted on three public benchmark datasets demonstrate CEMA's superior performance compared to existing methods.
万方数据
