RNN Encoder-Decoder for the inference of regular human mobility patterns

In this study, we proposed a deep learning model to infer the daily individual mobility pattern from static census data. Our work was inspired by Google Brain team work on machine learning system to automatically produce captions that accurately describe images using recurrent encoder-decoder model. They also use a convolutional neural network to exploit the strong spatially local correlation present in their structured data i.e. images. Unfortunately, survey data are generally heterogeneous with unknown local structure. Thus we have adapted their model using instead an appropriate mixed-variate version of restricted Boltzmann machine (MVRBM). This leads to estimation of daily regular mobility displacements in the form of variable length sequence given input individual attributes. The prime strength of our approach is that the resulting mobility flows inherit all the attributes contained in the input census which are typically missing in portable digital media data. Moreover, the model makes use of land use and point of interest data. Optimized in this way, the model is scalable to apply to other places conditioned with censuses availability. Finally, it has been validated and showed its efficiency in a real context.