Does Data Augmentation Distort User Intent? Frequency-aware Adaptive Contrastive Learning for Sequential Recommendation

Contrastive Learning (CL) has emerged as an effective paradigm for enhancing Sequential Recommendation (SR) by generating informative self-supervised signals through data augmentation. However, existing augmentation strategies often disregard the frequency distribution of items and users, leading to over-perturbation of rare but informative interactions, which damages user intent and harms recommendation quality. In this work, we propose FACL (Frequency-aware Adaptive Contrastive Learning), a novel yet lightweight framework for robust and intent-preserving sequential recommendation. FACL introduces a micro-level frequency-aware augmentation strategy that adaptively protects low-frequency items and subsequences during perturbation, and a macro-level frequency-aware reweighting scheme that assigns higher importance to sequences with rarer interactions in the contrastive loss. Together, these strategies strike a balance between diversity and fidelity in generated positive views. We instantiate FACL on top of mainstream SR models and evaluate it on five datasets. Experimental results show that FACL consistently outperforms state-of-the-art methods, improving the recommendation performance by 5.8 % in average while maintaining robustness to rare-item sequences.