Two-Stage Graph Contrastive Distillation Recommendation Model

Abstract

Recommendation systems frequently experience substantial performance degradation during distribution shifts. Existing single-stage methods cannot balance the training robustness and test adaptability. We propose a two-stage graph contrastive distillation (TS-GCD) recommendation model, and establish a collaborative optimization framework of “training robustness construction and testing adaptive fine-tuning.” In the training phase, distributionally robust optimization (DRO) is adopted to learn robust graph representations. In the testing phase, local-global dual views are constructed based on spectral domain frequency difference under different propagation depths, and adaptation is realized via self-distillation consistency constraint, thereby achieving optimization synergy between the two stages. Main experiments on popularity distribution shifts across four datasets verify the effectiveness of TS-GCD, and motivational analysis on other types of distribution shifts on two additional datasets are performed, providing references for future research.