Clustering driver’s destinations - using internal evaluation to adaptively set parameters

Abstract: With advanced navigation systems becoming ubiquitous in modern cars, the availability of detailed GPS data opens up new research areas in the fields of pattern analysis and data mining. By capturing the end-of-trip GPS points of each trip made by a driver, that driver’s meaningful destinations could be identified. The knowledge of these destinations can be used for route prediction, which in turn can be used for optimizing the motor control to decrease emissions. It can also be used for developing functions for autonomous vehicles. In this thesis a way of extracting these meaningful destinations from GPS data using clustering algorithms has been developed and evaluated. The result is a clustering procedure consisting of 2 steps of clustering. First a pre-clustering to divide the data into subsets corresponding to smaller spatial areas. Then, a refining clustering step for which the parameter of the algorithm is adapted to each subset. Adaptively setting the parameter for each subset is done by testing a set of parameters and evaluating the results internally, with the Silhouette coefficient, and choosing the parameter giving the best evaluation score. The best performing configuration of our procedure, according to our external evaluation method, is in par with the performance of DBSCAN with a supervised choice of parameter setting. Further evaluation of data sets from different areas of the world are needed to draw strong conclusions of the developed procedures performance.