A1692
Title: Constrained contextual bandits under systemic fluctuations
Authors: Yi Chen - Hong Kong University of Science and Technology (Hong Kong) [presenting]
Abstract: In online decision-making, unobserved systemic fluctuations often confound perceived outcomes, leading to a profit illusion where transient environmental surges trigger the premature exhaustion of scarce resources on suboptimal actions. Standard methods typically fail by conflating these global shocks with intrinsic treatment effects. To address this, a constrained contextual bandit model is formulated incorporating unobservable, time-varying intercepts to capture arbitrary exogenous volatility. An online framework is proposed that achieves statistical consistency by centralizing real-time data, a simple yet effective algebraic approach that decouples systemic noise from true signals without requiring evolution forecasting. By overcoming the complex temporal dependencies inherent in adaptive processes, sublinear regret bounds are rigorously established, demonstrating the policy's robustness. The method proves its efficacy in operational settings such as e-commerce coupon assignment and personalized medicine by preventing resource misallocation. Ultimately, this research establishes a critical design principle: explicitly modeling and isolating systemic fluctuations is essential for achieving operational resilience and long-term efficiency in volatile, budget-constrained environments.
