Authors: Daniel Fuller*, Memorial University of Newfoundland, Rui Zhang, University of Saskatchewan, Meghan Winters, Simon Fraser University, Yan Kestens, Universte de Montreal, Kevin G Stanley, University of Saskatchewan
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling, Geography and Urban Health
Keywords: spatial features, machine learning, global positioning systems, wearable devices, smartphones
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Congressional A, Omni, West
Presentation File: No File Uploaded
Purpose Collection of global positioning system (GPS) data from wearable devices, including smartphones, is increasingly common. Machine learning has considerable promise to better understand GPS data in many research areas. However, a standard feature set for GPS data has not been developed. The purpose of our work is to propose an early step in the development of a standard feature set for GPS data to distinguish and equate datasets with known demographic and geographic differences or similarities. Methods We used data from multiple GPS studies involving human participants, including the Saskatchewan Human Ethology Datasets (SHED) conducted in Saskatoon, Saskatchewan, and the Interventions, Research, and Action in Cities Team (INTERACT) studies, conducted in Victoria, Vancouver, Saskatoon, and Montreal. We build on our previous work and test newly developed features. We developed or applied three new features; convex hull with dwell time thresholds, scale independent entropy, and daily path mobility with multiple buffer sizes. We also tested three previously developed features; convex hull, entropy, and fractal dimension. We applied the features to the datasets and conducted statistical tests to determine the relationship between the features, whether the features could distinguish or equate datasets, and whether the features could discriminate different sociodemographic groups. Conclusions Our results suggest that our previous and newly developed features are able to distinguish and equate datasets and discriminate different sociodemographic groups. Interestingly, we show a strong correlation between the volume of the convex hull and the fractal dimension of the GPS traces.