Exploration vs. exploitation in coupon personalization

Abstract

Personalized offers aim to maximize profit by taking into account customer preferences inferred from past purchase behavior. For large retailers with extensive product offerings, learning customer preferences can be challenging due to relatively short purchase histories of most customers. To alleviate the dearth of data, we propose exploiting similarities among products and among customers to reduce problem dimensions. We also propose that retailers use personalized offers not only to maximize expected profit, but to actively learn their customers' preferences. An offer that does not maximize expected profit given current information may still provide valuable insights about customer preferences. This information enables more profitable coupon allocation and higher profits in the long run. In this thesis we 1) derive approximate inference algorithms to learn customer preferences from purchase data in real time, 2) formulate the retailers' offer allocation problem as a multi armed bandit and explore solution strategies.