Detecting Activity Types and Trip Purposes from Passive GPS Data: A Data Mining Approach - Space-Time Integration in Geography and GIScience - page 209

of placing much emphasis on data preprocessing, they introduce an innovative
approach called adaptive boosting to continuously train and improve the decision
tree, in which, each classifier votes for its predicted class, and the one who gets most
votes is considered to be the final class. 226 GPS trip records from 36 respondents
are collected to test the method and results show that the application of decision tree
with adaptive boosting can largely improve the accuracy of trip purpose detection.
It is argued, however, that decision tree is more suitable for cases with relatively
simple relationship with a small number of variables.
Chen et al. ( 2010 ) apply a probabilistic model (the multinomial logit model) to
determine trip purpose. The model estimates the probabilities of a trip having one
of several possible purposes from data on time of day, history dependence and land
use characteristics. Moiseeva et al. ( 2010 ) conduct a pilot experiment of applying an
artificial intelligent method with self-learning capability to detect transport modes
and trip purposes. They employ the Bayesian network that establishes probabilistic
relationships between input variables and outcomes. Transport modes and trip
purposes are classified based on the conditional probabilities estimated from input
variables on trip characteristics. The self-learning function is realized through
updating the conditional probabilities. A web-based prompted recall is presented
to respondents for confirming or correcting the derived activity-travel patterns,
which are used to update the conditional probabilities for subsequent classifications.
The results of an empirical application show that better accuracy of both mode
and purpose detection can be achieved through the self-learning mechanism. A
shortcoming of the Bayesian network is that it needs prior information from past
experiences to derive the probability function. Nevertheless, this study demonstrates
the potentials of artificial intelligent methods with self-learning for detecting trip
purposes. Thus more research efforts to explore the potentials of the artificial
intelligent methods for trip purpose detection are needed.

13.3 A Genetic Algorithm for Detecting Trip Purposes

To further explore the potentials of artificial intelligent methods for data mining
passive GPS data, this study proposes to apply the genetic algorithm, which is
characterized by strong capability to search for optimal solution through self-
learning, to detect trip purposes.
Genetic algorithm (GA) is an evolutionary type of algorithms (EA), which
provide effective ways to search for optimal solution to a problem through self-
learning (Buckles and Petry 1992 ). GA has been widely used in a great variety of
applications in the fields of science and engineering to solve classification-related
problems such as feature selection and classification (Punch et al. 1993 ), remote
sensing image classification (Yang et al. 2006 ), and classification and prediction
of precipitation occurrence (Sen and Öztopal 2001 ). In the past decade, there have
been a number of reported studies on GA applications in the field of transportation.
For instance, Gen et al. ( 1999 ) present a genetic algorithm to optimize large-scale