Online Learning Techniques for Pricing in e-Commerce Scenarios

The revenues of Online Travel Agencies (OTAs) come mainly from the profit they get reselling flight tickets. Therefore, the OTA's main goal consists in setting the optimal ticket price, i.e., the one which provides the maximum expected revenue from each flight ticket they sell. Since we are in a setting characterized by lack of information (uncertainty about the user preferences), non-stationarity (due to both competitors price oscillations and seasonal effects) and huge catalog (the profit for each flight should be considered separately), it is necessary to develop automatic techniques to handle the burden of the optimization procedure.

In this talk we formulate the pricing optimization problem as a contextual Multi-Armed Bandit (MAB) and present the design of novel MAB algorithms able to exploit the characteristics of pricing. In the first part, we consider typical assumptions about the pricing setting (i.e., monotonicity of the conversion rate, unimodality of the expected profit function and constraints on cost budget) to enhance the performance of the state of the art MAB algorithms. After that, we exploit the logged information available about past OTAs transactions to partition the decision space (i.e., the ticket catalog) and consequently speed up the learning process.