Not All Candidates are Created Equal: A Heterogeneity-Aware Approach to Pre-ranking in Recommender Systems

Researchers developed HAP (Heterogeneity-Aware Adaptive Pre-ranking), a new framework for recommender systems that addresses gradient conflicts in training by separating easy and hard samples. The system has been deployed in Toutiao's production environment for 9 months, achieving 0.4% improvement in user engagement without additional computational costs.

• →HAP framework solves gradient conflicts in pre-ranking by treating easy and hard training samples differently with dedicated optimization paths.
• →The system adaptively allocates computational resources, using lightweight models for all candidates and stronger models only for difficult cases.
• →Production deployment at Toutiao showed 0.4% improvement in app usage duration and 0.05% increase in active days over 9 months.
• →Traditional pre-ranking methods waste computation by uniformly scaling model complexity across heterogeneous samples.
• →Researchers released a large-scale industrial dataset to enable systematic study of candidate heterogeneity in pre-ranking systems.