RoleMix: Unified Architecture for Post-Click Conversion Rate Prediction
A recent publication on arXiv introduces RoleMix, an integrated interaction framework designed for predicting post-click conversion rates (PCVR) in industrial recommendation systems. RoleMix tackles the issue of structural discrepancies between sparse, unordered multi-field features and extensive, domain-specific behavioral histories by utilizing a shared, role-preserving token interface to represent both sequential and non-sequential data. It transforms non-sequential fields into clear semantic tokens that maintain user, item, pairwise, dense, contextual, and cross-feature roles. Extensive behavior domains are condensed into item- and context-aware sequence-query tokens through a two-stage hierarchical window attention mechanism. Stacked UniMixing-Lite blocks are employed to jointly refine global, semantic, and sequence-query tokens for PCVR prediction. The paper can be found on arXiv under ID 2607.22700.
Key facts
- RoleMix is a unified interaction architecture for PCVR prediction.
- It uses a shared, role-preserving token interface for sequential and non-sequential features.
- Non-sequential fields are converted into explicit semantic tokens.
- Long behavior domains are compressed via two-stage hierarchical window attention.
- Tokens are refined by stacked UniMixing-Lite blocks.
- The paper is on arXiv with ID 2607.22700.
- It addresses structural mismatch in industrial recommendation systems.
- The approach aims to improve cross-signal refinement.
Entities
Institutions
- arXiv