Contextual search in the presence of irrational agents
Abstract: Modern online marketplaces require decisions to be made sequentially. These decisions not only affect the system's performance on the current customer but may also have long-lasting effects, giving rise to a sequence of novel challenges. In this talk, I will focus on one example of such challenges: the need for robustness to data corruption and other model misspecifications. Classical machine learning approaches rely on collecting a batch of data and fitting a model to it -- this assumes that customers' behavior is identically and independently distributed. However, in practice, the behavioral models assumed are often slightly misspecified, e.g., due to the strategic behavior of participating entities or due to the fact that some agents do not subscribe to the assumed behavioral model. Motivated by this practical concern, I will will introduce an algorithmic framework for achieving robustness to such model misspecifications and will describe it in the context of a canonical revenue management setting, that of feature-based dynamic pricing .
The seminar will be mostly based on joint work with Akshay Krishnamurthy, Chara Podimata, and Robert Schapire. The corresponding paper can be found here. A preliminary version of this work appeared at STOC '21 and the paper is currently under journal submission.
