关键词: 序列推荐, 文本-ID融合, 混合专家系统, 多策略对比学习, 自适应标签平滑

Abstract: Significant progress has been made in sequential recommendation systems for capturing the evolution of user interests, where the effective use of side information plays a crucial role. For example, item-related textual data can be effectively fused with item IDs in a transformed feature space, and this text-ID semantic fusion can then be leveraged to enhance recommendation performance. However, existing models still suffer from limitations such as the complexity of user behaviors, limited generalization after heterogeneous information fusion, and data sparsity. This study proposes a text–ID sequential recommendation model named MCLALS (Multi-strategy Contrastive Learning and Adaptive Label Smoothing). The model first employed a domain-expert fusion network to process textual representations and dynamically integrate knowledge from different feature subspaces. Then, the processed text and ID representations were semantically fused at the sequence level through frequency-domain transformation to capture deep interactions among heterogeneous information. On the fused high-level representations, multi-strategy contrastive learning was applied to generate diverse augmented representations using multiple complementary data augmentation strategies, effectively alleviating data sparsity. Finally, adaptive label smoothing was adopted to optimize the training objective, enhancing the model’s robustness to noise and long-tail distributions. The MCLALS model was systematically evaluated on three public benchmark datasets: Food, Office, and OR. Experimental results show that the proposed model outperforms several baseline models, including SASRec（Self-attentive sequential recommendation） and TedRec（Text-ID Semantic Fusion for Sequential Recommendation）, in overall performance. In particular, on the OR dataset, the MCLALS model achieves at least a 6.69% improvement in NDCG@10 compared to the best baseline model, thereby verifying its effectiveness.

Key words: sequential recommendation, text-ID fusion, mixture-of-experts system, multi-strategy contrastive learning, adaptive label smoothing