NEWS
Contextual Bandits Research Accepted at ACM SIGIR 2025 Short Paper Track
We’re excited to share that our paper, “Counterfactual Model Selection in Contextual Bandits,” authored by Rakuten Institute of Technology (RIT), has been accepted to the short paper track at the 48th ACM SIGIR 2025, the premier conference on research and development in information retrieval!
Huge congratulations to authors, Shion Ishikawa, Young-joo Chung, Yun Ching Liu, and Yu Hirate for their hard work and dedication.
In this paper, we address the challenge of exploration efficiency in contextual bandit algorithms by introducing a novel counterfactual approach to model selection. We leverage unbiased Off-Policy Evaluation (OPE) to dynamically select base policies, resulting in algorithms (MetaEXP-OPE and MetaGreedy-OPE) that are more robust to model misspecification. Our experiments on synthetic and semi-synthetic data show that our methods significantly outperform existing policies like MetaEXP and MetaCORRAL.
We’re looking forward to presenting our research and connecting with the IR community in Padua, Italy, from July 13-17. You can find more information about SIGIR 2025 here: https://sigir2025.dei.unipd.it/
