A comparison of spatio-temporal prediction methods: a parking availability case study

In this paper, we present a comparative analysis of Statistical, Machine Learning and Deep Learning spatio-temporal models for parking occupancy prediction1. We evaluate such models on three public datasets, which are enriched by a set of hand-crafted features to take into account the temporal and spatial components when they are not natively handled by a model. Two approaches for this regression task are investigated: a univariate one and a multivariate one. In the former, we build a separate model for each parking lot. In the latter, a single model is used to predict the availability of all parking lots so as to learn the interactions and the co-movements among all time-series. All models exhibit similar performance. However, we highlight the higher effectiveness of gradient boosted methods when encompassing both temporal and spatial awareness in the feature space and of deep-learning models that take into account the spatial structure of the data.