Using Survival Prediction Techniques to Learn Consumer-Specific Reservation Price Distributions

Abstract

A consumer's "reservation price" (RP) is the highest price that s/he is willing to pay for one unit of a specified product or service. It is an essential concept in many applications, e.g., personalized pricing, auction and negotiation. While consumers will not volunteer their RPs, we may be able to predict these values, based on each consumer's specific information, using a model learned from earlier consumer transactions. This thesis proposes a novel framework of learning RP distributions that involves a model of formulating the relationship between consumers' RPs and their purchasing decisions, and a data collection method. Within this framework, we show a way to estimate the consumer-specific RP distribution using techniques from the survival prediction --- here viewing the consumers' purchasing choices as the censored observations. To validate our new framework of RP, we run experiments on realistic data, with four survival methods. These models performed very well (under three different criteria) on the task of estimating consumer-specific RP distributions, which shows that our RP framework can be effective. As we found that the multi-task logistic regression model (MTLR) dominated the other models under all three evaluation criteria, we explored ways to extend it, leading to extensions that are more general and more flexible. Moreover, we prove that it is the general regularizer, instead of the smoothness regularizer, that results in a smooth predicted distribution; this leads further simplification of the MTLR model.