Reproduce “Dual Conditional Diffusion Models for Sequential Recommendation” on OpenResearch

About this paper

Hongtao Huang, Chengkai Huang, Tong Yu, Xiaojun Chang, Wen Hu, Julian McAuley, Lina Yao

Researchers at the University of New South Wales and collaborators developed DCRec, a dual-conditional diffusion model that integrates both implicit and explicit user preference signals for sequential recommendation. The model achieved consistent and statistically significant performance improvements across multiple benchmark datasets, for instance, a 3.45% improvement in HR@5 and 3.16% in NDCG@10 on the Beauty dataset over strong diffusion baselines, while also improving computational efficiency during inference.

Key insights

• Jointly modeling both implicit (through full sequence noising) and explicit (through direct history guidance in reverse process) historical conditions consistently outperforms models relying on a single conditioning type.
• The proposed Conditional Layer Normalization and dual attention mechanism (self-attention followed by cross-attention with pure history) are critical for effectively integrating dual conditions and capturing nuanced sequential patterns.
• Adding noise to the entire concatenated sequence (history + target item) is more effective than adding noise only to the target item, particularly in scenarios with disparate sequence component lengths, due to better balance in noise application.

Summary by alphaXiv.