Revolutionizing Cascade Ranking with LCRON: End-to-End Training for Ads

Introduction

In large-scale Top‑K selection tasks such as recommendation and advertising platforms, cascade ranking has become the dominant architecture. Traditional training focuses on single‑stage optimization and ignores complex interactions between stages, leading to suboptimal system performance.

Challenges

Training objectives are misaligned with the overall goal of selecting ground‑truth items.

Lack of effective learning of inter‑stage cooperation.

LCRON Framework

Full‑space Sample Construction

LCRON builds training samples from the entire candidate space, providing rich feedback signals and ensuring the data distribution matches the system pipeline.

End‑to‑End Surrogate Loss

The framework formulates the problem as a survival‑probability estimation, using a differentiable surrogate loss that directly optimizes the probability that ground‑truth items survive through all cascade stages.

Key equations are represented using ... tags (omitted for brevity).

Auxiliary Loss for Tightening the Lower Bound

Additional stage‑specific auxiliary losses are introduced to shrink the gap between the surrogate loss lower bound and the true survival probability, encouraging consistent top‑k selections across stages.

Experiments

Public Benchmark

On the RecFlow dataset, LCRON outperforms baselines in joint recall while maintaining competitive single‑stage metrics.

Kuaishou Commercial Deployment

Deployed in Kuaishou’s advertising system, LCRON achieved significant improvements in revenue and conversion rates compared with FS‑LambdaLoss and ARF.

Business Impact and Outlook

LCRON has been fully deployed since January 2025, unifying recall and coarse‑ranking training, boosting ad revenue, and becoming the dominant traffic channel. Future work includes further system cost optimization and broader application of the end‑to‑end paradigm.