Mixture-of-tastes Models for Representing Users with Diverse Interests
This addresses a fundamental limitation in recommendation systems for users with multiple distinct interests, though it appears incremental as it builds on existing models.
The paper tackles the problem of recommendation systems treating user tastes as unimodal when users actually have diverse interests, showing that modeling tastes as a mixture leads to large increases in recommendation quality across both deep sequence-based and traditional factorization models.
Most existing recommendation approaches implicitly treat user tastes as unimodal, resulting in an average-of-tastes representations when multiple distinct interests are present. We show that appropriately modelling the multi-faceted nature of user tastes through a mixture-of-tastes model leads to large increases in recommendation quality. Our result holds both for deep sequence-based and traditional factorization models, and is robust to careful selection and tuning of baseline models. In sequence-based models, this improvement is achieved at a very modest cost in model complexity, making mixture-of-tastes models a straightforward improvement on existing baselines.