Dual-Oriented Contrast for Recommendation with A Stop-Gradient Operation
Byungkook Oh, Yul Kim, Bumky Min · 2023
Recently, contrastive loss is adopted as a main objective of recommender systems. InfoNCE-like losses penalize hard negative items more and control the strength of penalties with a temperature, called hardness-aware sensitivity. However, since they leverageuser->item patterns in a non-symmetric way, negative items are pushed away from anchor users and attract semantically-similar items to each other, focusing on the distribution of item embeddings. We point out that user embeddings also have inherent semantic structures that can be captured fromitem->user patterns. This paper presents Dual-oriented Contrast(DuCo), a novel symmetric learning objective for recommendation to learn more comprehensive representations fromusereftrightarrowitem patterns. DuCo controls user-/item-centric hardness-aware sensitivities and simultaneously optimizes the score distributions over sampled items (user-oriented contrast) and users (item-oriented contrast). This aims to explore ideal user and item distributions that are locally clustered and globally uniform. However, since user-/item-side temperatures are interdependent, naive control over temperatures may break the underlying semantic structures of the other side. To this end, we employ a stop-gradient operation to preserve the individual characteristics of user/item embedding distributions. Furthermore, we balance user-/item-oriented contrasts during learning to maintain consistent high-rank performance (e.g., recall@1). Empirical results show that DuCo contributes to the top-k user and item prediction simultaneously, and outperforms state-of-the-art learning objectives across different backbones from ID-based to neighbor-based encoders.