Research map / Bandits and online learning / Contextual bandits and Thompson sampling
Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback
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 |
|---|---|---|---|
| Use ARRoW-CB with |Λ|=8 instead of ARRoW-CB with |Λ|=2 (ablation) | ARRoW-CB · 524 OpenML classification datasets, aggregated noise conditions | normalized error | Helped |
| Adaptively select among a richer set of source weightings rather than use Sim-Bandit's weighting | ARRoW-CB · 524 OpenML classification datasets, aggregated noise conditions | normalized error | Helped |
| Use supervised warm-start data when costs are aligned | ARRoW-CB · Aligned costs, c^s = c^b | normalized error | Helped |
| Use adaptive weighting to combine moderately biased warm-start data with bandit feedback | ARRoW-CB · CYC noise p=0.25; CB ground truth | normalized error | Helped |
| Use adaptive weighting to combine majority-corrupted warm-start data with bandit feedback | ARRoW-CB · MAJ noise; CB ground truth | normalized error | Helped |
| Use misleading warm-start data with CYC noise p=1.0 instead of ignoring it | ARRoW-CB versus Bandit-Only · CYC noise p=1.0; CB ground truth | normalized error | Hurt |
| Increase ε-greedy exploration from 0.0125 to 0.1 (ablation) | ARRoW-CB · CYC noise p=1.0; CB ground truth | normalized error | Helped |
| Increase ε-greedy exploration when warm-start bias is high (ablation) | Contextual bandit algorithms · Adversarially corrupted warm-start data | normalized error | Mixed |
| Use supervised-only greedy predictions rather than exploration when costs are aligned | Sup-Only versus ARRoW-CB · Aligned costs, c^s = c^b | normalized error | Helped |
| Adaptively combine the sources across different warm-start ratios | ARRoW-CB · MAJ noise p=0.25; warm-start ratios 2.875 to 184 | normalized error | Helped |
| Weight supervised and bandit examples equally in a warm-started UCB strategy under source mismatch | Warm-started UCB · Two-armed bandit with misaligned deterministic costs | regret: Ω(Δ exp(Δ² n^s/16)) | Hurt |
| Use a larger set of candidate source weightings instead of selecting only between source-only options (ablation) | ARRoW-CB · Aggregated experimental conditions | normalized error | Helped |
Closest papers by meaning
- Leveraging Good Representations in Linear Contextual Bandits (ICML 2021 · similarity 0.90)
- Model Selection for Contextual Bandits (NeurIPS 2019 · similarity 0.90)
- BISTRO: An Efficient Relaxation-Based Method for Contextual Bandits (ICML 2016 · similarity 0.90)
- Active Context Selection Improves Simple Regret in Contextual Bandits (NeurIPS 2026 · similarity 0.90)
- Efficient Contextual Bandits with Uninformed Feedback Graphs (ICML 2024 · similarity 0.89)
- From Contextual Combinatorial Semi-Bandits to Bandit List Classification: Improved Sample Complexity with Sparse Rewards (NeurIPS 2025 · similarity 0.89)
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.