Applying semi-supervised learning method for cellphone-based travel mode classification

Transportation mode detection is important in understanding traffic conditions, facility performance, and residents' daily movements. Using GPS data collected from personal cellphones, this study analyzed mode classification methods for participants in Moscow, Idaho. Principal component analysis and semi-supervised Gaussian mixture models were implemented as major machine learning techniques applied in the classification task. For the study of two-mode classifiers, the prediction accuracy was found to be 65.71% and 88.00% for motorized and non-motorized trips, respectively. For the four-mode classifiers (bike, bus, drive, and walk), the model correctly predicted 66.67% and 57.14% of the trips for the Drive and Walk modes. The prediction accuracy for Bike and Bus was not as high due to the small number of trips observed in these two modes. Ultimately, the model built with PC scores performed better than model with non-transformed variables.