Tca4rec: contrastive learning with popularity-aware asymmetric augmentation for robust sequential recommendation

Abstract Sequential recommender systems play a pivotal role in modern recommendation scenarios by capturing users’ dynamic interests through their historical interactions. While existing methods often rely on sophisticated deep models to enhance recommendation quality, they suffer from performance degradation due to sparse supervision signals and popularity bias in the training data. In this paper, we propose TCA4Rec, a robust sequential recommendation framework that addresses these challenges via a novel two-stage contrastive learning approach. Our framework incorporates an additional memory module to aggregate sequence embeddings, thereby providing flexible and generalized representations of user preferences. To mitigate popularity bias, we derive an Asymmetric Multi-instance Noise Contrastive Estimation (AMINCE) loss function that supplies rich, bias-aware training signals, while our two-stage training strategy significantly reduces the over-dominance of popular items during optimization. Extensive experiments are conducted on three real-world datasets. The experimental results demonstrate that TCA4Rec achieves significant improvements over state-of-the-art baselines, in terms of recommendation accuracy and robustness against sparse and noisy data. It attains relatively gains of 19.26% in HR@5 and 17.97% in NDCG@5 on the Amazon-sports dataset. The code is available at https://github.com/shixiaoyu0216/TCA4Rec/tree/main .