ResGNN: a residual GNN approach for leveraging general user preferences in session-based recommender systems

In recent years, session-based recommender systems (SBRSs) have emerged as pioneers for intelligent recommendation environments by capturing short-term user preferences without requiring direct access to user history. However, the challenge remains in effectively considering both short-term and long-term user preferences. Graph neural networks (GNNs) have shown promise in this field by leveraging the structural information of user–item interactions, allowing for a comprehensive representation of short-term session dynamics and long-term preference patterns. Building upon these advancements, this paper presents a novel GNN approach for session-based recommendation leveraging general user preferences, referred to as ResGNN. Whereas existing GNN models experience a drop in performance when going through multiple propagation steps due to over-smoothing and the vanishing gradient problem, ResGNN addresses this by integrating residual connections within the GNN architecture. It captures complex item–item relationships within and across sessions while incorporating user preferences through multiple propagation steps. This research presents alternative learning techniques, specifically pre-training and live tuning, to effectively address rapid changes in user preferences within SBRSs. ResGNN is evaluated on two real-world datasets, demonstrating significant improvements compared to state-of-the-art SBRS methods. Notably, it shows enhanced ranking capabilities with a + 2.81% improvement in MRR@10 and + 1.15% in MRR@20 on the YooChoose dataset and a + 4.91% increase in MRR@10 and + 2.24% in MRR@20 on the Diginetica dataset.