ACRec: Adaptive Contrastive Learning for Sequential Recommendation Based on Transformer

Sequential recommendation aims to predict the next item a user will interact with based on past behaviors. Deep learning and contrastive learning have significantly improved recommendation accuracy. However, traditional contrastive learning methods require manually designed data augmentation strategies, such as item masking or reordering, which are dataset-dependent, time-consuming, and difficult to generalize. To address these limitations, we propose ACRec, a self-attentive contrastive learning framework based on the Transformer architecture. ACRec adaptively learns data characteristics through paired and augmented contrastive learning, enhancing recommendation performance. Extensive experiments on public datasets demonstrate that ACRec outperforms state-of-the-art sequential recommendation models such as CoSeRec and CBit. Specifically, compared to CBit, ACRec achieves an average improvement of +2.86% Recall and +4.46% NDCG on the Video Games dataset, +2.63% Recall and +1.66% NDCG on Automotive, and +1.5% Recall and +0.58% NDCG on Grocery. These results confirm the effectiveness of adaptive augmentation in contrastive learning for sequential recommendation.