Implicit GPS-based bicycle route choice model using clustering methods and a LSTM network

Biking is gaining in popularity all around the world as a healthy and environmentally friendly mode of transportation. Urban policies tend to encourage citizens to use bicycles. This can be done by creating new cycling infrastructures, the renovation of old ones or the deployment of bike-sharing systems (BSS). These policies having a cost, understanding and predicting the behavior of cyclists has become a necessity in order to optimize them. Classical methods analyzing cyclists' route choices use external factors and generated choice sets of paths along with a logit model to create a discrete route choice model. Nevertheless, few studies focus on the predictive capacity that this type of model can offer. In this paper, we developed a prediction-centered bicycle route choice model. Our model is created without using external factors or choice sets of paths as in the more classical methods. The idea of our method is to use deep and machine learning algorithms on GPS tracks. These algorithms learn representations from the data which replace explicit factors. To build the model, we clustered the GPS tracks using DBSCAN. The clusters allow to identify the cyclists' preferred road segments and are used to create paths using them. A method weighting the road graph weights is developed to create paths passing through the preferred road segments of a given cluster. A LSTM is finally trained in order to retrieve a cluster from a shortest path between an origin/destination pair. Tracks created by our model are more similar to the original GPS tracks than the shortest paths or tracks generated by a prominent path computation service.