Enhancing Recommendation System Consistency via Feedback Cascading and LTR Models

This paper addresses consistency challenges in multi-stage recommendation systems through feedback cascading and learning-to-rank (LTR) models. The study proposes a teacher-student architecture where fine-tuned models guide coarse sorting, improving alignment between stages. Key innovations include ΔnDCG-based loss functions and bid-aware ranking optimizations. Experimental results show 4.8% RPM increase during A/B testing.

The work tackles two core issues: sparse data in coarse sorting (SSB problem) and ranking inconsistency between modules. By leveraging end-to-end sequence learning and constraint-based optimization, the approach achieves better score alignment with fine-tuned models. The final solution combines pointwise loss constraints with sequence-aware weighting to maintain both ranking consistency and bid responsiveness.