Fusing Sequential Behavior and Item Attributes: Sequential Recommendation using Dual-Stream Fusion Transformer

Sequential recommendation systems are a key technology for capturing users‘ dynamic preferences, but they often face the fundamental limitation of data sparsity. To alleviate this problem, recent contrastive learning-based models, such as CaDiRec, have achieved high performance by using diffusion models for context-aware data augmentation. However, these approaches are limited as they only learn from the sequential context of items, failing to leverage the rich semantic context derived from unique item attributes. We propose AF-CaDiRec, a model that enhances item representations by effectively fusing item ID embeddings with item attribute embeddings using a Dual-Stream Fusion Transformer. Experimental results on the MovieLens-1M dataset showed that AF-CaDiRec improved performance over the baseline CaDiRec. Specifically, it achieved gains of 1.36% in HR@10 and 1.85% in NDCG@10. For K=5, NDCG@5 improved by 1.23% while HR@5 remained at nearly the same level. This signifies that AF-CaDiRec maintains the ability to include correct items in the top list while improving the ranking quality by placing them higher. These results demonstrate that fusing attribute information is effective in enhancing the overall quality of the recommendation list.