Reproduce “Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale” on OpenResearch

About this paper

Yongsen Pan, Yuxin Chen, Zheng Hu, Xu Yuan, Daoyuan Wang, Yuting Yin, Songhao Ni, Hongyang Wang and 3 more

AdaSID, an adaptive framework for learning Semantic IDs (SIDs), improves multimodal recommendation quality by dynamically regulating collisions in the discrete ID space. It distinguishes between harmful ambiguities and semantically admissible sharing, while adaptively allocating regulation pressure both spatially and temporally. The framework achieved superior offline recommendation performance and demonstrated significant online gains in industrial A/B tests, including a 0.98% increase in Gross Merchandise Volume (GMV).

Key insights

• Distinguishing between harmful collision and semantically admissible sharing is crucial for effective SID learning, preventing over-separation and preserving useful semantic relationships.
• Dynamically adjusting collision regulation based on spatial congestion and training progress leads to a more robust and discriminative discrete SID space.
• A holistic approach that balances discrete structure formation (early training) with recommendation-oriented alignment (later training) is more effective than static weighting of objectives.

Summary by alphaXiv.