Sequential Recommendation with Diverse Supervised Contrastive Views

Zitong Zhu, Meixiu Long, Junfa Lin, Jiahai Wang

Abstract: Recently, contrastive learning has been used in sequential recommendation to address data sparsity issue. Among various contrastive learning methods, supervised contrastive learning has shown excellent performance by utilizing supervised contrastive views. However, current supervised contrastive views in sequential recommendation are sampled from existing sequences in a dataset, resulting in insufficient supervised contrastive information and low-quality contrastive views. To address the problem, this paper proposes sequential Recommendation with Diverse supervised contrastive views (DivRec). DivRec has two main components, contrastive views generation and contrastive views integration. Contrastive views generation constructs diverse supervised contrastive views from a generation perspective, where insertion augmentation and non-autoregressive generation are utilized to generate locally-diverse and globally-diverse supervised contrastive views, respectively. Contrastive views integration integrates multiple contrastive views for their complementary effects, where a leader-follower strategy in the integration component helps construct attribute-compatible contrastive pairs to mitigate conflict issues. Additionally, diffusion mechanism is introduced as a noise augmentation method to further improve the diversity of contrastive views. Extensive experiments demonstrate the effectiveness and the state-of-the-art performance of DivRec. The code and appendix are available at https://github.com/zzzzzdev/DivRec.

External IDs:dblp:conf/adma/ZhuLLW24