Contrastive Learning for Sequential Recommendation With Negative Selection and Augmentation
Renqi Jia, Xiaokun Zhang, Weitao Xu, Chen Ma · Artificial Intelligence for Engineering · 2025
ABSTRACT Contrastive learning (CL) alleviates data sparsity in sequential recommendation (SR) by leveraging positive‐negative sample discrimination. However, most existing CL methods focus on positive sample augmentation and treat all nonpositive samples in a batch uniformly as negatives, ignoring the varying quality and limited quantity of negatives. Through empirical analysis, we observe that CL‐based SR methods using hard negative samples—those highly similar to the anchor—achieve significantly better performance, as they force the model to learn finer‐grained feature distinctions. Motivated by this insight, we propose a novel method, termed contrastive learning for sequential Recommendation with Negative selection and Augmentation (NARec), which focuses on the selection and augmentation of negative samples. To select high‐quality negative samples, the selection module identifies hard negative samples from both in‐batch and global sample pools. To enrich the quantity of the negative samples, the augmentation module generates high‐quality negative samples in a learnable manner to expand the sample pool. By integrating a multi‐task learning framework, NARec jointly optimises the recommendation, contrastive learning and data augmentation tasks. Extensive experiments on multiple public datasets demonstrate NARec's superior recommendation performance compared to competitive baselines. Further studies confirm the pivotal role of the negative quality in enhancing the recommendation performance.