Research map / Learning theory and algorithms / Prediction and decision losses
Learning Cascade Ranking as One Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LCRON, which trains a full cascade ranking system (retrieval + ranking) as one network by optimizing a lower bound on the end-to-end survival probability of ground-truth items plus a per-stage auxiliary Recall loss, combined via uncertainty weighting.
Problem
Training cascade ranking systems stage by stage misaligns objectives with the end-to-end recall of ground-truth items and fails to learn collaboration between stages.
Approach
Derives a differentiable lower bound (product of per-stage top-q selection probabilities via NeuralSort soft permutation matrices) on the joint survival probability of ground-truth items, optimizes it end-to-end with a cross-entropy loss, and adds a single-stage Recall loss over the full soft permutation matrix to tighten the bound and supply supervision when survival probabilities are near 0.
Key results
- On RecFlow Period 1 last-day test, LCRON reaches joint Recall@10@20 of 0.8732±0.0005 vs. 0.8678 for ARF-v2 and 0.8666 for FS-LambdaLoss.
- In a 15-day Kuaishou advertising A/B test against FS-LambdaLoss, LCRON gave +4.1% revenue and +1.6% ad conversions (ARF-v2: +1.66% and +0.65%).
- On a three-stage RecFlow setting, LCRON reaches joint Recall@10@20 of 0.7390 vs. 0.7332 for ARF-v2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train the cascade jointly with LCRON losses instead of ARF-v2 | DSSM retrieval + DIN ranking · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8678±0.0009 -> 0.8732±0.0005 | Helped |
| train the cascade jointly with LCRON losses instead of ARF-v2 | DSSM + MLP + DIN · RecFlow Period 1 last-day test, three-stage cascade | Joint Recall@10@20: 0.7332±0.0020 -> 0.7390±0.0008 | Helped |
| deploy LCRON instead of FS-LambdaLoss | Kuaishou advertising system · 15-day online A/B test | Revenue: +4.1% | Helped |
| deploy LCRON instead of FS-LambdaLoss | Kuaishou advertising system · 15-day online A/B test | Ad Conversions: +1.6% | Helped |
| add L_e2e to LCRON (ablation) | LCRON · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8710 ± 0.0013 -> 0.8732 ± 0.0005 | Helped |
| add L_single to LCRON (ablation) | LCRON · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8712 ± 0.0004 -> 0.8732 ± 0.0005 | Helped |
| replace ARF's relaxed loss with L_single | ARF · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8608±0.0006 -> 0.8678±0.0009 | Helped |
| use NeuralSort with LCRON instead of standalone NeuralSort (ablation) | NeuralSort ranking on RecFlow · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8210±0.0016 -> 0.8732±0.0005 | Helped |
| use SoftSort with LCRON instead of standalone SoftSort (ablation) | SoftSort ranking on RecFlow · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8103±0.0013 -> 0.8723±0.0008 | Helped |
| vary L_single weight with L_e2e fixed to 1 (ablation) | LCRON · RecFlow Period 1 last-day test, two-stage cascade | Joint Recall@10@20: 0.8712; 0.8706; 0.8705; 0.8723; 0.8731; 0.8730; 0.8720; 0.8733 | Mixed |
Surprising findings
- Retrieval-stage NDCG@10 of LCRON is not the best; BCE obtains the highest retrieval Recall@10@30 (0.9706 vs. 0.9700 for LCRON), and the paper notes LCRON does not dominate individual stage metrics.
- Standalone differentiable sorting operators (NeuralSort, SoftSort) trained directly on the labels perform far worse than the LCRON loss (0.8210 and 0.8103 vs. 0.8732 joint recall).
Limitations
- Setting q1 and q2 to 10 during training differs from actual system quotas (30 and 20), trading off gradient stability against interpretability; larger q may cause gradient conflicts.
- Lower bound may be loose; tightness depends on consistency between stages, and UWL does not guarantee optimal weighting.
- Experiments limited to two-stage industrial setting and a single public benchmark (RecFlow); industrial three-stage not evaluated.
Experimental setup
Datasets: RecFlow
Benchmarks: RecFlow
Metrics: Recall@10@20 NDCG@10 Recall@10@30 advertising revenue ad conversions
Baselines: BCE ICC RankFlow FS-RankNet FS-LambdaLoss ARF ARF-v2
Builds on: ARF FS-LTR ICC RankFlow NeuralSort
Closest papers by meaning
- Both Supply and Precision: Sample Debias and Ranking Consistency Joint Learning for Large Scale Pre-Ranking System (AAAI 2025 · similarity 0.84)
- FinalMLP: An Enhanced Two-Stream MLP Model for CTR Prediction (AAAI 2023 · similarity 0.83)
- From Pairwise to Ranking: Climbing the Ladder to Ideal Collaborative Filtering with Pseudo-Ranking (AAAI 2025 · similarity 0.83)
- GoalRank: Group-Relative Optimization for a Large Ranking Model (ICLR 2026 · similarity 0.83)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.83)
- No More Tuning: Prioritized Multi-Task Learning with Lagrangian Differential Multiplier Methods (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.