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

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Prevailing pre-ranking methods that indiscriminately mix heterogeneous training samples suffer from gradient conflicts, where hard samples dominate training while easy ones remain underutilized, leading to suboptimal performance.technical 90%
• The common practice of uniformly scaling model complexity across all samples in pre-ranking is inefficient, as it overspends computation on easy cases and slows training without proportional gains.technical 90%
• The Heterogeneity-Aware Adaptive Pre-ranking (HAP) framework mitigates gradient conflicts through conflict-sensitive sampling coupled with tailored loss design, while adaptively allocating computational budgets across candidates.method 90%
• HAP disentangles easy and hard samples, directing each subset along dedicated optimization paths, applying lightweight models to all candidates and stronger models only on hard ones.method 90%
• Deployment of HAP in the Toutiao production system for 9 months yielded up to a 0.4% improvement in user app usage duration and a 0.05% improvement in active days, without additional computational cost.result 95%
• The methodology might be heavily tailored to Toutiao's existing infrastructure, and adapting it to other platforms may require significant customization.limitation 85%
• The improvements from HAP, while important at scale, may be perceived as marginal on smaller platforms.limitation 80%
• The paper releases a large-scale industrial hybrid-sample dataset to enable the systematic study of source-driven candidate heterogeneity in pre-ranking.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card