Machine Learning from imbalanced data-sets: an application to the bike-sharing inventory problem

One of the major issue bike sharing operators struggle to deal with is the bicycle rebalancing activity, i.e. optimizing the fleet location reducing the related activity cost. In order to reduce operational cost generated by rebalancing and to facilitate the adoption of bike sharing by users, it is extremely important to estimate the correct value of bicycles (and available docks in case of station-based bike sharing), that is the optimal inventory level. In this paper we investigate the potential of using machine learning techniques for estimating the inventory level to address the station-based bike sharing static rebalancing in the case of imbalanced data-set. Specifically, Random Forest (RF) and Gradient Tree Boosting classifiers have been proposed, together with a new iterative approach based on RF. All the methods have been tested adopting real world data of New York City bikes together with weather data.