A data mining approach to rapidly learning traveler activity patterns for mobile applications.

摘要

The swift growth in the number of GPS devices has led to a boom in the number of mobile applications attempting to exploit this rapidly growing market. As a result, understanding travelers and their information needs has become a major topic of interest. While many studies have examined learning traveler behavior, they have primarily concentrated on the destination and route information. There are two key weaknesses of these studies. First, they require a lengthy history of the person be collected before a reasonable model can be built. Second, they focus on the travel itself rather than the reason for the travel. While trip information is useful, the reason for the travel likely is more useful to mobile applications aimed at influencing the user's plans. The purpose of this study is to address both of these points: reducing learning time and examining the reason for the travel rather than just the trip itself.;To accomplish these goals, this work examines using an interdisciplinary approach to combine transportation planning activity-based modeling methods with data mining techniques to learn individual patterns. This work demonstrates that such a model can be tailored to the patterns of an individual traveler, allowing projections of their future trips, activities, and planning flexibility to be made. Second, due to the abstraction of the model, an extensive history of the user is not necessary to build a reasonable model of the traveler. Traditionally, however, this type of model has required collecting a detailed activity history that is likely more burdensome than most mobile application users would accept.;This research addresses this challenge by creating an activity model of a traveler while greatly reducing the data entry required by the user. The primary contribution of this work is a set of techniques for quickly learning the travel activity patterns of individuals with limited user interaction. This is achieved through three main areas: (1) leveraging passive data to augment user entered data; (2) introducing techniques to reduce the impact of missing data on prediction quality; and (3) supplementing user patterns with general patterns from other sources.