Dual-view co-contrastive learning for multi-behavior recommendation

Multi-behavior recommender systems (MBR) typically utilize multi-typed user interactive behaviors (e.g., purchase, click, view and add-to-cart) in learning user preference on target behavior (i.e., purchase). Existing MBR models have a natural deficiency, namely the sparse supervision signal problem, which may degrade their actual recommendation performance to some extent. Inspired by the recent success of contrastive learning in mining additional supervision signals from raw data itself, in this work, we seek to exploit the co-contrastive learning to enhance multi-behavior recommendation. However, the following two key challenges remain to be addressed: (1) How to select proper views in multi-behavior modeling; (2) How to design a difficult but effective contrastive task. To tackle the above challenges, we devise a novel co-contrastive learning framework without sampling named D ual-view C o-contrastive L earning (DCL). Unlike traditional contrastive learning methods that generate two augmented views by corruption, we construct two views from different aspects of user preference. Technically, we first leverage the multi-behavior interaction graph to enhance two views that capture both local collaborative signals and high-order semantic information simultaneously. And then two asymmetric graph encoders are performed on both views, which recursively exploit the different structural information to generate ground-truth samples to collaboratively supervise each other by co-contrastive learning and finally high-level node embeddings are learned. Moreover, the view complementary mechanism and divergence constraint further make the two view encoders different but also complementary, so as to improve co-contrastive learning performance. Extensive experiments on two real-world datasets indicate that MBR can be significantly augmented under the regime of co-contrastive learning and then achieves the state-of-the-art performance.