Research map / Label-efficient and robust learning / OOD and anomaly detection
AdaO2B: Adaptive Online to Batch Conversion for Out-of-Distribution Generalization
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 AdaO2B, an adaptive online-to-batch conversion that combines a bandit algorithm's sequence of policies with context-aware MLP weights, with OOD generalization bounds, evaluated on synthetic and KuaiRec data.
Problem
Classic online-to-batch conversion assumes i.i.d. data, but streaming recommender data drifts over time, so fixed conversion mechanisms generalize poorly to out-of-distribution testing data.
Approach
Combine the sequence of policies learned by a contextual bandit algorithm into a batch learner using combination weights produced by a context-aware MLP, trained with MSE on observed rewards from selected data buffers (sliding window, reservoir, or data-dependent selection), with a theoretical OOD error bound tied to weighted regret and JS/KL divergence.
Key results
- On synthetic data, AdaO2B-S with SBUCB reached 0.7848 average reward, a 6.23% gain over the best baseline (LL, 0.7388).
- On KuaiRec, AdaO2B outperformed the best baselines by 1.94% (SBUCB), 2.41% (EXP3-B), and 2.33% (BLTS-B).
- AdaO2B with EXP3-B on synthetic data improved over the best baseline by 14.58% using the data-dependent selection variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdaO2B with sliding-window data selection instead of the strongest baseline | SBUCB · Synthetic | average reward: +6.23% | Helped |
| Use AdaO2B with data-dependent selection instead of the strongest baseline | EXP3-B · Synthetic | average reward: +14.58% | Helped |
| Use AdaO2B with sliding-window data selection instead of the strongest baseline | BLTS-B · Synthetic | average reward: +4.23% | Helped |
| Use AdaO2B with reservoir-sampling data selection instead of the strongest baseline | SBUCB · KuaiRec | average reward: +1.94% | Helped |
| Use AdaO2B with sliding-window or reservoir-sampling selection instead of the strongest baseline | EXP3-B · KuaiRec | average reward: +2.41% | Helped |
| Use AdaO2B with reservoir-sampling data selection instead of the strongest baseline | BLTS-B · KuaiRec | average reward: +2.33% | Helped |
| Use sliding-window selection rather than reservoir sampling (ablation) | AdaO2B · Synthetic | average reward | Helped |
| Use data-dependent selection rather than reservoir sampling (ablation) | AdaO2B · Synthetic | average reward | Helped |
| Use reservoir sampling rather than the other data-selection approaches (ablation) | AdaO2B · KuaiRec | average reward | Helped |
Surprising findings
- AdaO2B outperformed the fully online learning oracle (FOL) on synthetic data, which the authors attribute to exploration hurting accuracy in the synthetic test environment.
- On KuaiRec, FOL with EXP3-B had lower average reward than with other bandit backbones, attributed to EXP3's adversarial exploration term.
Limitations
- Theoretical guarantees assume linear (or convex) expected rewards and bounded weighted regret.
- Evaluation uses a simulated online environment built from an MF model rather than live feedback.
- Experiments are limited to synthetic data and a single real-world dataset (KuaiRec).
Experimental setup
Datasets: synthetic Gaussian contexts KuaiRec
Benchmarks: synthetic OOD recommendation simulation KuaiRec
Metrics: average reward
Baselines: Best Batch Learner Last Learner Simple Averaging Conversion Voting Conversion Constant Weight Fully Online Learning
Builds on: online to batch conversion sequential batch UCB EXP3 Thompson sampling for bandits
Closest papers by meaning
- Online Boosting Adaptive Learning under Concept Drift for Multistream Classification (AAAI 2024 · similarity 0.84)
- Diversified Batch Selection for Training Acceleration (ICML 2024 · similarity 0.83)
- OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? (ICML 2025 · similarity 0.82)
- GREATS: Online Selection of High-Quality Data for LLM Training in Every Iteration (NeurIPS 2024 · similarity 0.82)
- Efficient Non-stationary Online Learning by Wavelets with Applications to Online Distribution Shift Adaptation (ICML 2024 · similarity 0.82)
- A Simple and Adaptive Learning Rate for FTRL in Online Learning with Minimax Regret of $\Theta(T^{2/3})$ and its Application to Best-of-Both-Worlds (NeurIPS 2024 · 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.