Download: PDF.
“A Data Mining Approach To Rapidly Learning Traveler Activity Patterns For Mobile Applications” by Chad A. Williams. Ph.D. Dissertation, Department of Computer Science, University of Illinois at Chicago, April 2010.
Recent work has begun to examine predicting traveler patterns at an individual level through two main approaches. Transportation planners have examined an activity based analysis approach in part due to the theoretic transferability across people with similar characteristics. While these models are transferrable, the predictions are more for a type of individual than a specific individual and require a considerable amount of data from the traveler. Mobile and ubiquitous computing researchers, on the other hand, have focused on learning the patterns of individuals based on passive location observations through GPS traces of that specific individual. While the predictions of this type of model are tailored to the individual, a lengthy history of the individual's travels must be collected before significant patterns and adequate coverage emerge. This work introduces a way to combine these two approaches to learn an individual's activity patterns with limited data input required from the traveler.
This study examines how activity patterns of individuals may be quickly learned with limited traveler input. Reducing participant burden is a critical component of making this type of learning practical for applications such as an intelligent traveler's assistant (ITA). An analysis is presented of how technologies such as GPS can be combined with data mining methods to greatly reduce the effort needed to develop an activity model of the traveler. A primary contribution of this work was developing a technique to model the activity and travel patterns of an individual on an ongoing basis that greatly reduces data collection requirements. This advance opens up new possibilities for knowing current context and projected context for applications like an ITA without overwhelming the user with data entry requirements. While this research is aimed at enhancing real-time travel applications, there are a number of other applications of this work as well. An example of this is reducing user burden in travel surveys; a design of an adaptive travel survey is introduced as one way this type of approach can be used to improve common tasks such as activity surveys. The benefits of this approach are illustrated in the reduction of participant burden in a multi-day activity based travel survey.
Due to the challenge of reducing data requirements while still developing a meaningful model of the traveler, new methods were needed. To address this goal, techniques that would allow ongoing data collection to reduce the questions asked of the traveler were examined. One way this was accomplished was through a new procedure that allowed models of the individual to be continually adapted as new data was collected. Second, a new mining algorithm was introduced that further reduced data requirements by improving predictions despite missing values. This technique proved particularly helpful in allowing non-intrusive collection methods such as GPS traces to largely replace the need for many questions as the collection period progressed. Finally, a technique was developed to use data from other surveys to improve the activity model of the individual despite differences in the city in which the other surveys were collected. The combination of all of these advances is shown to significantly reduce data requirements in an ongoing collection effort. To demonstrate the benefits of these advances, a real-world example of how this technique might be used to improve existing collection efforts was developed in the form of a multi-day activity survey. Results based on 42 households demonstrate the advantages of this work in being able to reduce the respondent burden with limited impact on the results of the survey.
In summary, this research extends current ideas and introduces algorithms and techniques for learning individual travel behavior. The focus of this research was to leverage the transferrable aspects of travel behavior and patterns to reduce learning time, while also creating a richer model of the individual traveler. This research effort introduces algorithms and techniques needed to address the problem of learning and predicting the activity needs of an individual for anticipating their associated travel demands. The results presented demonstrate that the techniques introduced in this work can make a significant impact on curtailing the impact on participants for multi-day activity learning. Advances such as these are likely to make the proposition of longer surveys or the interaction required for real-time traveler applications like an ITA much more practical.
Keywords: GPS, travel behavior, activity patterns, data mining, machine learning, Learning Activity Patterns of Individuals, activity modeling, patterns of individuals
BibTeX entry:
@phdthesis{W10,
author = {Chad A. Williams},
title = {A Data Mining Approach To Rapidly Learning Traveler Activity Patterns For Mobile Applications},
school = {Department of Computer Science, University of Illinois at
Chicago},
type = {{Ph.D.} Dissertation},
month = {April},
year = {2010}
}