Noise-reducing graph neural network with intent-target co-action for session-based recommendation

Abstract

Session-based recommendation (SBR) originates from a real-world need to provide effective recommendation solutions for unlogged users. How to utilize short interaction sequences of anonymous users for practical recommendations has become a critical issue. Current research on SBR generally suffers from noisy terms interfering with prediction, difficulty in capturing long-term dependencies, and ineffective utilization of session-level information. Inspired by the outstanding learning ability of the sparse attention mechanism for distant information and reduced noise through sparse transformations, we propose an SBR model. We combine the graph neural network with sparse graph attention mechanisms to learn enhanced and denoised target information and long-term preferences. In addition, we propose a similar-intent collaboration module that can efficiently utilize session-level collaboration information. Combining the information, the model generates a reliable representation of the session. Our model is superior to the current SOTA model on three widely used datasets (Diginetica, Tmall, RetailRocket). Compared with the best baseline, we improved by 5.09%, 3.86%, and 5.45% with P@20 and by 12.63%, 22.78%, and 3.06% with MRR@20.