mlsanityData to 14 Oct 2026

Choosing the Better Bandit Algorithm under Data Sharing: When Do A/B Experiments Work?

arXiv 2507.11891 ↗First posted 16 Jul 2025stat.ML cs.LG math.STcited by 4

Not on the map. Why?

Abstract (arXiv)

We study A/B experiments that are designed to compare the performance of two recommendation algorithms. Prior work has observed that the stable unit treatment value assumption (SUTVA) often does not hold in large-scale recommendation systems, and hence the estimate for the global treatment effect (GTE) is biased. Specifically, units under the treatment and control algorithms contribute to a shared pool of data that subsequently train both algorithms, resulting in interference between the two groups. In this paper, we investigate when such interference may affect our decision making on which algorithm is better. We formalize this insight under a multi-armed bandit framework and theoretically characterize when the sign of the difference-in-means estimator of the GTE under data sharing aligns with or contradicts the sign of the true GTE. Our analysis identifies the level of exploration versus exploitation as a key determinant of how data sharing impacts decision making, and we propose a detection procedure based on ramp-up experiments to signal incorrect algorithm comparison in practice.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Builds on 39 of its 59 references that are in the map. First posted on arXiv 16 Jul 2025.

Builds on

39 of 59 references.

References · 39

  1. 2019Introduction to Multi-Armed Banditscited by 481
  2. 2016Corralling a Band of Bandit Algorithmscited by 100
  3. 2019Confidence Intervals for Policy Evaluation in Adaptive Experimentscited by 95
  4. 2020Inference for Batched Banditscited by 63
  5. 2019Model selection for contextual banditscited by 60
  6. 2019Social Learning in Multi Agent Multi Armed Banditscited by 48
  7. 2014Design and analysis of experiments in networks: Reducing bias from interferencecited by 42
  8. 2021Off-Policy Evaluation via Adaptive Weighting with Data from Contextual Banditscited by 40
  9. 2020Experimental Design in Two-Sided Platforms: An Analysis of Biascited by 39
  10. 2020Design and Analysis of Switchback Experimentscited by 38
  11. 2019OSOM: A simultaneously optimal algorithm for multi-armed and linear contextual banditscited by 28
  12. 2022A Simple and Provably Efficient Algorithm for Asynchronous Federated Contextual Linear Banditscited by 28
27 more
  1. 2021Interference, Bias, and Variance in Two-Sided Marketplace Experimentation: Guidance for Platformscited by 22
  2. 2022Markovian Interference in Experimentscited by 22
  3. 2021Multiple Randomization Designs: Estimation and Inference with Interferencecited by 20
  4. 2022A survey on multi-player banditscited by 19
  5. 2022Competition, Alignment, and Equilibria in Digital Marketplacescited by 17
  6. 2022Switchback Experiments under Geometric Mixingcited by 17
  7. 2020Adaptive Experimental Design with Temporal Interference: A Maximum Likelihood Approachcited by 15
  8. 2020Randomized Graph Cluster Randomizationcited by 14
  9. 2020Multitask Bandit Learning Through Heterogeneous Feedback Aggregationcited by 14
  10. 2020Reducing Interference Bias in Online Marketplace Pricing Experimentscited by 13
  11. 2020Competing Bandits: The Perils of Exploration Under Competitioncited by 12
  12. 2024Data-Driven Switchback Experiments: Theoretical Tradeoffs and Empirical Bayes Designscited by 12
  13. 2021A/B Testing for Recommender Systems in a Two-sided Marketplacecited by 11
  14. 2021Design and Analysis of Bipartite Experiments under a Linear Exposure-Response Modelcited by 9
  15. 2023An Experimental Design for Anytime-Valid Causal Inference on Multi-Armed Banditscited by 9
  16. 2021Learning from an Exploring Demonstrator: Optimal Reward Estimation for Banditscited by 8
  17. 2023Reducing Symbiosis Bias Through Better A/B Tests of Recommendation Algorithmscited by 8
  18. 2022Exploiting Neighborhood Interference with Low Order Interactions under Unit Randomized Designcited by 7
  19. 2023Evaluating Online Bandit Exploration In Large-Scale Recommender Systemcited by 7
  20. 2023Experimenting under Stochastic Congestioncited by 6
  21. 2025Double Machine Learning for Causal Inference under Shared-State Interferencecited by 4
  22. 2023Producer-Side Experiments Based on Counterfactual Interleaving Designs for Online Recommender Systemscited by 3
  23. 2023Tackling Interference Induced by Data Training Loops in A/B Tests: A Weighted Training Approachcited by 3
  24. 2024When Does Interference Matter? Decision-Making in Platform Experimentscited by 3
  25. 2024Asymptotically and Minimax Optimal Regret Bounds for Multi-Armed Bandits with Abstentioncited by 1
  26. 2024Individual Regret in Cooperative Stochastic Multi-Armed Banditscited by 1
  27. 2025Two-Sided Prioritized Ranking: A Coherency-Preserving Design for Marketplace Experimentscited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026Design Experiments to Compare Multi-armed Bandit Algorithmscited by 1
  2. 2026AI-Generated Measurements for Identification and Inference with Missing Data: A Weak Shadow Variable Approachcited by 0
  3. 2026Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharingcited by 0
  4. 2026BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Testscited by 0