Research map / Bandits and online learning / Contextual bandits and Thompson sampling
A Contextual Bandit Bake-off
Alberto Bietti, Alekh Agarwal, John Langford
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use IWR reduction instead of IPS (ablation) | Bag · 516 OpenML datasets, optimized hyperparameters | significant win-loss difference: +55 | Improved |
| use IWR reduction instead of DR (ablation) | Bag · 516 OpenML datasets | significant win-loss difference: +25 | Improved |
| use IWR reduction instead of IPS (ablation) | epsilon-greedy · 516 OpenML datasets | significant win-loss difference: +128 | Improved |
| use IWR reduction instead of DR (ablation) | epsilon-greedy · 516 OpenML datasets | significant win-loss difference: +150 | Improved |
| use -1/0 loss encoding instead of 0/1 (all datasets) (ablation) | Greedy · multiclass/multilabel datasets | significant wins/losses of -1/0 vs 0/1: 132 / 42 | Improved |
| use -1/0 loss encoding instead of 0/1 (all datasets) (ablation) | epsilon-greedy · multiclass/multilabel datasets | significant wins/losses: 94 / 27 | Improved |
| use -1/0 instead of 0/1 on datasets with at least 10,000 examples (ablation) | RegCB-optimistic · 64 large datasets | significant wins/losses: 10 / 18 | Hurt |
| use -1/0 instead of 0/1 on datasets with at least 10,000 examples (ablation) | Cover-NU · 64 large datasets | significant wins/losses: 14 / 20 | Hurt |
| shift RCV1 costs from [0,1] to [-1,0] (ablation) | all methods · RCV1 | progressive validation loss | No clear effect |
| shift MSLR/Yahoo costs with c=-1 (ablation) | Greedy · MSLR | progressive validation loss: 0.798 -> 0.791 | Improved |
| shift Yahoo costs with c=-1 (ablation) | Greedy · Yahoo | progressive validation loss: 0.593 -> 0.589 | Improved |
| optimize hyperparameters per dataset instead of fixing them (ablation) | Cover-NU vs RegCB-optimistic · all datasets (521) | significant win-loss difference: RO beats C-nu by 28 -> C-nu beats RO by 33 | Mixed |
Closest papers by meaning
- Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback (ICML 2019 · similarity 0.93)
- Beyond UCB: Optimal and Efficient Contextual Bandits with Regression Oracles (ICML 2020 · similarity 0.92)
- BISTRO: An Efficient Relaxation-Based Method for Contextual Bandits (ICML 2016 · similarity 0.91)
- Contextual Bandits with Large Action Spaces: Made Practical (ICML 2022 · similarity 0.91)
- Active Context Selection Improves Simple Regret in Contextual Bandits (NeurIPS 2026 · similarity 0.91)
- Efficient Contextual Bandits with Uninformed Feedback Graphs (ICML 2024 · similarity 0.90)
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.
