RoleMix: Unifying Sequential and Non-Sequential Features via Semantic Tokenization for Post-Click Conversion Rate Prediction

RoleMix introduces a unified token interface for post-click conversion rate prediction, addressing the mismatch between sparse, unordered multi-field features and long behavioral histories. It converts non-sequential fields into role-preserving semantic tokens covering user, item, pairwise, dense, contextual, and cross-feature evidence. Long behavior domains are compressed into item- and context-aware sequence-query tokens using two-stage hierarchical window attention. Global, semantic, and sequence-query tokens are jointly refined by UniMixing-Lite blocks. On the KDD Cup 2026 Tencent UniRec Challenge, RoleMix reports 83.648% online AUC, 1.953% above the official industrial baseline.