Enhancing Multi-behavior Sequential Recommendation via Transformer-Based Cross-Layer Contrastive Learning

Multi-behavior sequential recommendation methods have surpassed the limitations of single-behavior sequence modeling by capturing diverse user interactions. However, existing approaches often rely on shallow, static feature extraction across behavior types, overlooking the rich dynamic semantics embedded in user interaction sequences. Furthermore, traditional data augmentation strategies are susceptible to random noise during auxiliary view construction, resulting in semantic incompleteness and representation sensitivity. To address these limitations, we propose A Multi-behavior Sequential Recommendation Framework with Transformer and Cross-layer Contrastive Learning (MST-CCL). Specifically, we first construct a heterogeneous item-behavior graph using users’ multi-type interactions and contextual information. Next, we encode this graph via a graph attention network, where cross-layer contrastive learning simultaneously enforces noise robustness and semantic consistency alignment. Finally, a self-attention mechanism is employed to adaptively fuse users’ temporal dynamics with multi-behavioral item relationships. Experimental results demonstrate that MST-CCL significantly outperforms state-of-the-art baselines on the Taobao and IJCAI datasets. Notably, our method achieves substantial improvements in Hit Rate (HR@10), with gains of +2.60% on Taobao and +2.10% on IJCAI, validating the effectiveness of cross-behavior signal mining and multi-view semantic alignment via contrastive learning.