Reproduce “Beyond Negative Transfer: Disentangled Preference-Guided Diffusion for Cross-Domain Sequential Recommendation” on OpenResearch

About this paper

Xiaoxin Ye, Chengkai Huang, Hongtao Huang, Lina Yao

DPG-Diff, a diffusion-based model, addresses challenges in cross-domain sequential recommendation by disentangling user preferences into domain-invariant and domain-specific components. The model demonstrates superior performance across real-world datasets, outperforming prior state-of-the-art models like DREAM by up to 28.2% in Hit@10, while also showing enhanced robustness to noisy interaction data.

Key insights

• Decomposing user preferences into distinct domain-invariant and domain-specific components is essential for effectively leveraging cross-domain knowledge while mitigating negative transfer.
• Diffusion models, with their iterative denoising framework, are well-suited for learning robust representations from complex and noisy sequential user behaviors in a generative manner.
• Guiding the diffusion process with explicitly disentangled preferences enables the model to selectively transfer knowledge across domains, promoting more accurate and noise-resilient recommendations.

Summary by alphaXiv.