Travel Mode Choice Prediction Using Deep Neural Networks with Entity Embeddings

Abstract

The prediction of travel mode preference, like many other choice prediction problems, may depend on categorical features of the choice options or the choice makers. Such categorical features need to be meaningfully encoded for better modeling and understanding. Problem-invariant encoding representations of the categorical features, such as one-hot encoding or label encoding, can severely limit the power of prediction models. We propose deep neural networks with entity embeddings for travel mode choice prediction. We adopt the entity embedding technique to jointly learn meaningful representation of categorical variables and accurate travel mode predictions. Experiments using the London travel dataset show that deep neural networks with entity embedding technique outperform neural networks with other encoding techniques, as well as tree-based models. Besides, we found that the learned embeddings can boost the performances of tree-based models by substituting categorical features with the neural network learned features. Finally, we verify that entity embedding can learn meaningful representations of the categorical features using feature visualization at low dimensional space.