Reproduce “Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation” on OpenResearch

About this paper

Guoqing Hu, An Zhang. Shuchang Liu, Wenyu Mao, Jiancan Wu, Xun Yang, Xiang Li, Lantao Hu, Han Li and 2 more

PreferGrow introduces a discrete diffusion model for recommendation systems that directly models user preference ratios via a "preference fading" mechanism, addressing the mismatch between continuous diffusion processes and discrete user interactions. The model consistently outperformed various state-of-the-art recommenders across five benchmark datasets and improved personalization through non-preference user modeling.

Key insights

• The use of an idempotent fading matrix for the forward process ensures both the Markov property and reversibility of the discrete diffusion, providing theoretical guarantees for the reverse generation.
• Modeling preference ratios directly via a fading mechanism offers a nuanced approach to capturing preferences, allowing for flexible integration of point-wise, pair-wise, hybrid, and adaptive negative sampling strategies.
• The score entropy loss, adapted for discrete diffusion, is proven to share the same descent direction and optima as the Binary Cross-Entropy loss, validating its suitability for ranking-oriented recommendation tasks.

Summary by alphaXiv.