List Recommendation via Co-attentive User Preference Fine-Tuning

Beibei Li, Beihong Jin, Xinzhou Dong, Wei Zhuo

Abstract: Most existing recommendation models focus on individual items. However, in many real-world scenarios, items are organized into lists (also called bundles) composed of multiple items with certain common correlations, e.g., video lists containing videos in the same genres, which makes list recommendation a significant task. In this paper, we propose a model named CAPLE which learns coarse-grained and fine-grained user preferences from user-list and user-item interactions respectively to achieve list recommendation. In particular, CAPLE develops a co-attention module, which fine-tunes user preferences by learning collaborative relations between user preferences of different granularities via jointly modeling tangled interactions of two different granularities. We conduct extensive experiments on real-world datasets. The performance enhancement on recall@K and NDCG@K verifies the effectiveness of the model.