International Journal of Computing and Digital Systems | 2026
Authors: Djenane M.A.; Boudaa B.; Abouaissa A.; Brahmia M.E.A.
DOI: 10.12785/ijcds/1571112706
Journal: International Journal of Computing and Digital Systems
Year: 2026
Publisher: University of Bahrain
Document Type: Article
Open Access: All Open Access; Gold Open Access
Cited by: 0
The advent of session-based recommender systems (SBRS) has significantly enabled recommendations to users without considering their historical data. Recent graph neural network (GNN) approaches have further advanced SBRS by enabling the integration of general user preferences (i.e., long-term preferences or global interests) directly from current session interactions, eliminating the need for historical user data. These approaches extract general user preferences by combining all session item embeddings with the final user-item interaction (i.e., click) as the user’s current interest. However, assuming the last interaction always represents the current user interest is limiting, as previous interactions may better capture true interests. This paper introduces a novel voting strategy that identifies the most valuable interaction within each session through a learnable mechanism with temperature-controlled softmax, enabling a more accurate representation of user interest. Unlike existing methods that rely solely on the final click, our approach amalgamates three complementary signals, which include global session embedding, the identified valuable interaction, and the last user-item interaction. These signals are then processed through an attention mechanism for integrating refined general user preferences. Experiments conducted on the YooChoose 1/64 and Diginetica datasets demonstrate consistent improvements over state-of-the-art GNN-based methods. Specifically, the Hit Rate at 20 (HR@20) increased by +1.15% and +1.60%, while the Mean Reciprocal Rank at 20 (MRR@20) improved by +0.29% and +3.63%, respectively. Optimal performance was achieved with a temperature setting of 0.1 in the voting layer. These results confirm that effectively identifying valuable interactions beyond the final click significantly enhances the accuracy of session-based recommendations. © 2026, University of Bahrain. All rights reserved.
General User Preferences; Graph Neural Networks; Session-Based Recommender System; Voting Strategy