Contextual search in the presence of irrational agents
Abstract
Contextual search lies at the heart of multiple applications such as feature-based dynamic pricing and personalized medicine. Recent work has made great advancements on how to effectively learn in contextual search settings even when the contexts (product features in pricing or clinical data of the patients) arrive in non-i.i.d. patterns. However, a strong assumption on these models is that the valuation of the agents is a linear combination of the user's contexts and a ground truth that is consistent across rounds and fully determines their behavior (full rationality assumption). In practice, this can be violated both because the models may have misspecifications and because the agents may occasionally behave in ways inconsistent with what the modes prescribe (or irrationally). However, even with small amounts of such noise (adversarial or stochastic), most current algorithms completely fail.
In this work, I will propose a unified way to tackle contextual search at the presence of model misspecifications by introducing a framework that captures different behavioral models and loss functions. I will then present an algorithm that provides guarantees which gracefully degrade with the extent of "irrationality" that the agents display and discuss the key algorithmic and analytical ideas. Finally, I will offer a broader view on thinking about hybrid models in sequential decision-making that combine both a) adaptivity to benign structures such as full rationality and b) robustness to misspecifications such as the ones arising due to irrationality.
Paper information: The talk is based on joint work with Akshay Krishnamurthy (MSR NYC), Chara Podimata (Harvard), and Robert Schapire (MSR NYC). A preliminary version with some of the technical results can be found here: https://arxiv.org/pdf/2002.11650.pdf. An updated version with a more unified perspective on behavioral models, loss functions, and algorithms for the setting will be available in a few weeks. The paper is currently under major revision in Operations Research.
Bio
Thodoris Lykouris is a postdoctoral researcher in the machine learning group of Microsoft Research NYC. His research focus is on data-driven sequential decision-making and spans across the disciplines of machine learning, operations research, theoretical computer science, and economics. He completed his Ph.D. in 2019 from Cornell University where he was advised by Éva Tardos. During his Ph.D. years, his research has been generously supported by a Google Ph.D. Fellowship and a Cornell University Fellowship. He was also a finalist in the INFORMS Nicholson and Applied Probability Society best student paper competitions.
