QCMP-CL: New contrastive learning method boosts sequential recommendations with quality-aware weighting
Learns diverse augmentations and adaptively weights them to reduce false positives.
Existing contrastive learning (CL) methods for sequential recommendation often rely on heuristic data augmentations (e.g., item cropping or masking) that can remove crucial items or break transition patterns, leading to semantic drift. To address this, Wei Wang proposes Quality-Aware Collaborative Multi-Positive Contrastive Learning (QCMP-CL), a framework that uses learnable collaborative augmentations instead. The method creates two augmented views: one based on sequences leading to the same target item (same-target context) and another from sequences with similar item embeddings (semantic similarity). This dual approach enhances diversity while preserving user intent.
A key innovation is the quality-aware mechanism, integrated directly into the model's representations. It estimates each view's reliability based on the confidence of its augmentation operations—high-confidence views contribute more to the contrastive loss, while low-confidence ones are downweighted. This adaptive weighting mitigates the false-positive issue common in CL. Evaluations on three real-world datasets (e.g., Yelp, Amazon) show QCMP-CL consistently outperforms existing CL-based baselines, demonstrating the effectiveness of combining diverse learnable augmentations with quality-aware supervision.
- QCMP-CL uses learnable augmentation to generate two complementary views: same-target sequences and semantically similar sequences, avoiding heuristic drift
- A quality-aware mechanism adaptively weights each view's contribution based on augmentation confidence, reducing false positives
- Outperforms state-of-the-art CL-based sequential recommenders on three real-world datasets (e.g., Yelp, Amazon)
Why It Matters
Improves recommendation accuracy in e-commerce and streaming by preserving user intent and reducing noisy signals in contrastive learning.