Characterization and Prediction of Car Sharing usage exploiting Points of Interest information

Recently, free-floating car-sharing systems were introduced as a novel way of mobility allowing users to reserve a car shortly before the rental with their smartphone. This approach is more dynamic than previous station-based car sharing systems, because users can start their trip wherever a car is available and terminate it anywhere in the operator’s area. Hence, this new form of car sharing is not only appealing to users with the need for a car but also to those who want to speed up their travel. One drawback of this system is that the user does not know in advance if there will be a car available nearby. This work solves this problem by predicting future car availabilities around Points of Interest (POIs) in the city. We train Random Forest models on features from the car sharing and POI datasets in order to generate predictions for the car availability in the near future. By applying this method to car sharing systems in Portland, Seattle and Turin, we are able to outperform the baseline of predicting the last recorded value. Therefore, our predictions can be used by users to inform themselves about future car availability or by car sharing operators to relocate their fleet. In addition, we discover behavioral patterns of users by applying sequence pattern mining to the car sharing data. Results from Portland and Turin show that we can extract different sequences with respect to both, specific POIs and POI categories. Moreover, this work shows that the extracted sequences coincide with sequences from a check-in database, where the users explicitly specify that they are visiting a POI at a given timestamp. Therefore, the discovered sequences of car sharing can be used to study the general movement of citizens. Furthermore, a temporal and spatial contextualization was conducted. Dividing the time of the day into four timeslots, differences in the frequency and confidence of the discovered sequences are found in each of the timeslots. Similarly, we split the area of the car sharing operator into multiple smaller areas and study the sequences extracted for each area. The results suggest that car sharing usage patterns highly depend on the area, as well as the time of the day. The discovered sequences can be used to study origins and destinations of car sharing trips, as well as the general movement of citizens. For instance, this can help the car sharing operator to decide which areas provide a good market to expand to and to discover general mobility patterns.