Mode choice models have been used widely to forecast the relative probabilities of using available travel modes. These depend on mode-related and traveler-related characteristics. On the other hand, smartphones are increasingly being used to collect sensors’ data relating to trips made after selection of a suitable mode. Such sensors’ data may be correlated with decision-making process of travelers regarding travel mode selection. Discrete Choice Modelling is used to simulate this decision-making process by computing utilities of various travel alternatives, and then calculating their respective probabilities of being selected. In this paper, multinomial logit (MNL) mode choice model is utilized to enhance the prediction capacity of supervised learning algorithm i.e. Weighted Random Forest. To make the procedure less energy-intensive, GPS data was used only to locate the origin and destination of any trip, to be incorporated in mode choice model. Afterwards only accelerometer data was utilized in feature selection for the learning algorithm. One tenth of the classified data was used to train the algorithm whereas rest was used to test it. Results suggested that with incorporation of MNL, the overall prediction accuracy of learning algorithm was increased from 93.75% to 99.08%.
The psychological decision-making process model of non-commuting travel mode choice under parking constraints
