Bridging Sequential and Contextual Features with a Dual-View of Fine-grained Core-Behaviors and Global Interest-Distribution

A research team led by Yi Xu has introduced CDNet (Core-Behaviors and Distributional-Compensation Dual-View Interaction Network), a novel architecture designed to improve click-through rate (CTR) prediction in recommendation systems. Traditional models face a fundamental trade-off: they either compress user behavior sequences into a single vector (losing important details) or analyze every individual action (which is computationally expensive and noisy). CDNet bridges this gap by examining user behavior from two complementary perspectives simultaneously.

The first view performs fine-grained analysis, identifying and focusing on the specific user behaviors most relevant to the candidate item. This captures precise, actionable signals. The second view operates at a coarse-grained level, modeling the user's overall interest distribution against the item's context to provide a holistic understanding. By combining these views, CDNet effectively models the interplay between sequential user actions and contextual item features without imposing a significant computational burden.

Extensive experiments validate CDNet's effectiveness, showing it outperforms existing CTR prediction models. The dual-view architecture allows it to capture nuanced behavioral details that single-vector approaches miss, while avoiding the noise and expense of analyzing every action individually. This makes CDNet particularly valuable for real-world applications like e-commerce and content platforms where accurate, efficient recommendations directly impact user engagement and revenue.