Title: Single-Index Contextual Bandits: Learning and Inference
Speaker: Sakshi Arya, Assistant Professor, Department of Mathematics, Applied Mathematics, and Statistics, Case Western Reserve University
Abstract: This talk focuses on statistical inference for contextual bandits with structured reward models. In particular, it considers single-index bandits, where the effect of covariates on each action is mediated through a low-dimensional projection, balancing interpretability and flexibility. While single-index structure is well studied in classical regression, bringing it to online decision-making introduces new challenges for uncertainty quantification because data are collected adaptively and observations are dependent. The talk outlines a methodological framework for learning these models sequentially while attaching principled uncertainty summaries to the learned components. Emphasis is placed on the theoretical issues posed by adaptive data collection and on the types of inferential guarantees that can be achieved, with brief examples illustrating how uncertainty summaries support interpretability in practice.