Application of machine learning in air-ticket pricing in China

Abstract

This dissertation predicts the air ticket price to provide advice on the immediate or postponed purchase of a trip. The proposed methodology is based on the statistical learning of a price evolution model from the joint information of the trip attributes. The main originality consists in representing the price evolution with the inhomogeneous punctual process of the price variation. This representation was used to group flights from the Qunar.com database into similar behaviors and to build a common model for each of the identified behaviors. We then implemented a learning method to model the price evolution. This model provides a predictor for the occurrence of a price drop over a given period and therefore offers advice on the immediate or postponed purchase of a trip to the customer. The research is organized in three main phases. First, we introduced a new method of representing time series of prices. This representation compared the series with each other and applied data mining algorithms. The approach is based on a preliminary statistical modeling for the dynamics of time price series and the theory of point processes. We first transformed the time series by point processes. Then we modeled these series of returns by an estimate of the intensity in gray level. In a second step, we proposed a segmentation of learning data in order to extract standard behaviors using unsupervised learning techniques. Based on our new representation of time series, we applied segmentation algorithms and extracted average behaviors called centroids. With each centroid, we simulated price curves for a behavioral prediction. In a third phase, we applied supervised learning algorithms on the attributes of the flights in each group to allocate new flights to a centroid with their attributes. The first chapter describes the data structure and explains the relevance of the decision support module. The next three chapters analyze the process of developing the representation, the data segmentation, and the learning phase with different settings for each of the steps. The experimental results are present with a comparative study of the different configurations for the algorithms and the approaches.