Deriving Rules for Trip Purpose Identification from GPS Travel Survey Data and Land Use Data: A Machine Learning Approach

Abstract

Passive GPS travel survey has proven an innovative alternative to the traditional paper-and-pencil method, as it can record exact time and location of travel activities without incurring much burden on survey respondents. Lying in the core of this technique is how to derive accurate travel information (e.g., trip purpose) from GPS data. Previous studies rely on simple rules that are manually extracted from the inherent structure of GPS data streams. These methods, however, are difficult to generalize for other applications and lack helpful hints for new research. This paper presents a machine learning approach to deriving trip purpose from GPS track data coupled with other relevant data sources. This approach employs a number of attributes (i.e., time stamp and land-use type of trip ends, a set of spatiotemporal indices of travel, and demographic and socioeconomic characteristics of respondents) to construct a decision tree for purpose of classification. Each attribute provides partial evidence to the depiction of a given purpose, but this depiction may or may not be monotonic, and none of them can work alone toward the goal. A reasoning procedure using the adaptive boosting technique was designed to explore how these attributes could work together to achieve trip purpose derivation. This technique generated multiple decision trees to improve the classification results through a mechanism of voting from these trees. Each tree was constructed in the depth-first fashion with the root node and the split of subsequent nodes being determined on the basis of the gain-ratio computed for the relevant attributes. This procedure was implemented in the C5.0 machine learning environment with 226 GPS trip records collected from 36 respondents. The experimental results seemed rather promising: using 10 iterations for adaptive boosting, an overall classification accuracy of 87.6% was achieved.