Research map / Learning theory and algorithms / Prediction and decision losses
Which Tricks are Important for Learning to Rank?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add stochastic smoothing to YetiLoss weights (vs. no smoothing) (ablation) | YetiLoss · Web10K, Yahoo S2, Istella-S (NDCG@10, MRR, MAP) | NDCG@10 / MRR / MAP: e.g. Web10K NDCG@10 48.49 -> 50.76 | Helped |
| Gaussian instead of logistic smoothing distribution (ablation) | YetiLoss · Web10K, Yahoo S2, Istella-S | NDCG@10 / MRR / MAP: differences within about 0.4 points | No clear effect |
| Add stochastic smoothing to YetiRank weights (vs. no smoothing) (ablation) | YetiRank · Web10K, Yahoo S2, Istella (NDCG@10, MRR, MAP) | NDCG@10: Web10K 49.05 -> 50.75 | Helped |
| Use YetiLoss (re-weighted gradients for target metric) instead of YetiRank | YetiRank · Web10K, Web30K, Yahoo S1/S2, Istella, Istella-S (MAP) | MAP: e.g. Istella 80.01 -> 82.66 | Helped |
| Use YetiLoss instead of YetiRank | YetiRank · Web10K, Istella, Istella-S (MRR) | MRR: Web10K 84.21 (YetiRank) vs 84.17 (YetiLoss); Istella 96.70 vs 97.16 | Mixed |
| Use YetiLoss instead of YetiRank | YetiRank · NDCG@10 benchmarks | NDCG@10: Istella 73.02 vs 71.60 for YetiLoss | Hurt |
| Use YetiRank instead of LambdaMART | LambdaMART · NDCG@10 tuned | NDCG@10: Web10K 49.38 -> 50.75 | Helped |
| Vary number of neighbors |p_i-p_j| in YetiLoss (k=1 vs 2 vs 3 vs all) (ablation) | YetiLoss · Web10K, Yahoo, Istella (MRR, MAP, ERR, NDCG@10) | NDCG@10 / MRR / MAP / ERR: no stable significant dependence; k=2 a good option | No clear effect |
| Replace YetiLoss with StochasticRank (direct optimization of non-convex smoothed loss) | StochasticRank · Web10K, Web30K, Yahoo S1, Yahoo S2, Istella, Istella-S | NDCG@10 / MRR / MAP / ERR: e.g. Istella NDCG@10 68.46 vs YetiLoss 71.60 | Hurt |
| Use CatBoost QueryRMSE baseline instead of ranking-specific losses | QueryRMSE · Web10K NDCG@10 tuned | NDCG@10: 50.55 vs YetiRank 50.75 | No clear effect |
Closest papers by meaning
- PiRank: Scalable Learning To Rank via Differentiable Sorting (NeurIPS 2021 · similarity 0.87)
- StochasticRank: Global Optimization of Scale-Free Discrete Functions (ICML 2020 · similarity 0.86)
- Decision-Focused Learning: Through the Lens of Learning to Rank (ICML 2022 · similarity 0.86)
- Are Neural Rankers still Outperformed by Gradient Boosted Decision Trees? (ICLR 2021 · similarity 0.86)
- Extreme Learning to Rank via Low Rank Assumption (ICML 2018 · similarity 0.86)
- RankSHAP: Shapley Value Based Feature Attributions for Learning to Rank (ICLR 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.