[1I5-GS-4b] Web intelligence

6:25 PM - 6:40 PM JST(9:25 AM - 9:40 AM UTC)

[1I5-GS-4b-04] An Interpretable and Controllable Multi-Interest Recommender Model

〇Keijiro Konishi¹, Kazushi Okamoto¹, Koki Karube¹, Kei Harada¹, Atsushi Shibata¹ (1. The University of Electro-Communications)

Keywords:

Recommender System,Interpretable,Controllable,Sequential Recommendation,Multi-Interest

Multi-interest recommendation systems effectively capture diverse user preferences by representing them with multiple vectors. However, the semantic meaning of the learned interest vectors is often unclear, therefore, making it difficult for users to explicitly control recommendation results based on their interpretation. This study proposes a model that enables explicit user control over recommendations while improving vector interpretability. Specifically, we impose an auxiliary task of item category prediction on each interest vector, in addition to the conventional next-item prediction task. Experimental results show that, despite a trade-off in prediction accuracy compared to baselines, our model achieves performance comparable to or better than existing approaches. Notably, the proposed method achieved a maximum 1.47-fold improvement in controllability when recommending items from specified categories, demonstrating enhanced controllability. Furthermore, qualitative analysis of the learned vectors indicates that our model assigns intuitive and consistent semantics to each interest vector.