Predicting trip purpose in a smartphone-based travel survey

Abstract

Population-wide travel surveys are conducted to investigate individuals’ patterns of traveling. These surveys are often burdensome. The widespread use of smartphones allows for collecting trip data without relying on traditional travel survey diary responses. Location monitoring data can be utilized to split the day into segments where someone is traveling (track) or stationary (stop). The current practice in official statistics is that respondents have to label each trip and stop, what mode of transport they travel with, and what the purpose of a stop is (e.g., shopping, work, school, home). By integrating smartphone GPS data with administrative, temporal, and spatial data, this paper studies to what degree it is possible to predict the purpose of a trip automatically. Multiple machine-learning models were trained and evaluated to unveil the effectiveness of stop-purpose prediction. In late 2022, Statistics Netherlands collected GPS data that contained 12 distinct labels denoting the purposes of the trips. The most optimal artificial neural network model and extreme gradient boosting technique obtained a balanced accuracy of 90% for the purpose of being at home. Primarily, classes that included only a small number of observations were erroneously categorized as classes with a large number of observations. Administrative data do not help to improve model prediction beyond spatiotemporal covariates. Increasing the duration of data gathering substantially enhanced the precision of the model. To summarize, smartphone-based travel data has considerable potential as a data source for trip purpose prediction but cannot yet be used to predict trip purpose automatically.