UniCon: a unified context-centric modeling paradigm for industrial CTR prediction

_reachsumit · x · 2026-09-04

UniCon proposes a unified context-centric architecture for CTR prediction that treats each display list as the basic modeling unit, organizing history and prediction targets as homogeneous context units.

Intra-context attention captures local coupling among items within a context, while inter-context attention models the dynamics of decision states across contexts. The authors argue existing token-level unified modeling stems from legacy feature engineering and misaligns with the user's decision process, limiting scaling efficiency in context-rich settings like e-commerce shelves and waterfall feeds.

Original post →

More from Research

Research channel →