LAUSR

Mobile telematics is a relatively new innovation that involves collecting data on driving behavior using the internal sensors in a smartphone rather than from an in-vehicle data recorder. However, telematics data are usually not labeled, which makes extracting driving patterns from them very difficult. Therefore, unsupervised learning algorithms play an important role in this field. In addition, most current research is based on datasets developed in a laboratory or from site investigations and questionnaires, which are very different from real-world driving behaviors. To advance unsupervised learning techniques in this field, and to fill the gap in findings based on real-world data, we have developed an unsupervised pattern recognition framework for mobile telematics data. The framework comprises three main components: a self-organizing map, a nine-layers deep auto-encoder, and partitive clustering algorithms. The SOM algorithm reduces the complexity of the data, the deep auto-encoder extracts the features, and the clustering algorithm groups driving events with similar patterns into behaviors. Further, given clustering with mobile telematics data is an under-researched area, we undertook an empirical comparison of five well-known clustering algorithms to determine the strengths and weaknesses of each method and which is best suited to categorizing driving styles. The study was conducted with a real-world insurance dataset containing 500,000 journeys by 2500 drivers, and the results were evaluated against three measures– Davis Boulding, Calinski Harabasz, and execution time. Overall, we find that k-means clustering and a self-organizing map were able to extract more accurate patterns than others. A statistical analysis of the 29 clusters produced by SOM and k-means, revealed 29 unique driving styles, all of which can be found in the transportation literature. The results from the study, with support from the corresponding literature review, demonstrate the efficacy of the presented framework in unsupervised settings. Additionally, the results provide a basis for developing a future risk analysis and automatic decision support system for usage-based insurance companies.