Origin-destination matrix estimation by deep learning using maps with New York case study

Origin-destination (OD) matrix estimation has conventionally required costly surveys or other forms of sensor data. In this paper, we propose that map data can supplement or partially replace such data if integrated in a deep learning method. Road network information and point of interest (POI) data are used for predicting interzonal trips within a city. The proposed learning approach is a hybrid model, combining a graph convolutional approach at the lower level with an upper level multilayer perceptron. We apply our method for estimating taxi trips in Manhattan and Brooklyn, New York, and obtain promising results for expanding an OD matrix with limited observations.