P3C: A Curiosity-Driven Recommendation Method Based on Preference and Consensus
Jia-Run Du, Ke Xu, Qian Gao · 2024
Existing recommendation systems mainly focus on achieving high accuracy, and ignore the psychological motivation behind human exploration. Inspired by the theory of curiosity, several curiosity-driven recommendation methods have been proposed. Despite the impressive success, these methods still suffer from two major limitations. One is that curiosity is only used to reorder the generated recommendation lists while ignoring its crucial role in the entire user decision-making process. Another is to evaluate conflict stimuli solely based on social friends factor, neglecting individual preference factor. In these regards, we propose a novel Preference-Consensus-Conflict Curiosity (P3C) method. As its core, a dual-branch network based on individual preference and social consensus is designed to model users’ conflict-driven curiosity. Furthermore, a re-weighting strategy combined with the curiosity is adopted for model training. Benefit from this, P3C can provide each user with a candidate item list that match their curiosity level. Extensive experiments conducted on MovieLens-1M and LastFM-2K demonstrate that P3C outperforms all the other competitors in diversity metrics (Coverage and Novelty), and achieves satisfactory performance in accuracy metrics (Precision, Recall, and NDCG). The code is available at https://github.com/Run542968/P3C.