Dual Correction Strategy for Ranking Distillation in Top-N Recommender System

• DOI: 10.1145/3459637.3482093
• Publication Date: 2021-11-15T10:31:19Z
• Authors: Youngjune Lee; Kee-Eung Kim
• Keywords: Learning to Rank

Knowledge Distillation (KD), which transfers the knowledge of a well-trained large model (teacher) to small (student), has become an important area research for practical deployment recommender systems. Recently, Relaxed Ranking (RRD) shown that distilling ranking information in recommendation list significantly improves performance. However, method still limitations 1) it does not fully utilize prediction errors student model, makes training efficient, and 2) only distills user-side information, provides insufficient view under sparse implicit feedback. This paper presents Dual Correction strategy (DCD), from teacher more efficient manner. Most importantly, DCD uses discrepancy between predictions decide be distilled. By doing so, essentially learning guidance tailored "correcting" what failed accurately predict. process is applied transferring as well item-side address user Our experiments show proposed outperforms state-of-the-art baselines, ablation studies validate effectiveness each component.