Applying transfer learning and various ANN architectures to predict transportation mode choice in Amsterdam

For long, statistical models have been used for transportation mode choice analysis, due to their ability to extract economic information from the model parameters. Recently, the application of Artificial Neural Nets to predict transportation mode choice is gaining ground, partly due to efforts that have led to an improved interpretability of this class of models. In this development, various innovations have been suggested concerning Neural Net architecture and hyperparameter tuning. Building on this, this paper investigates 3 similar Neural Net architectures to be applied to data from an Amsterdam case study. This data has been collected in 3 waves. Between the first and second collection period, the public transportation network in Amsterdam changed. A transfer learning approach is suggested to improve models that were trained on a single wave of data. Based on the test loss of the models from the transfer learning experiments, we conclude that this is a promising technique to use in this context, since it has shown to improve model performance.