From Cluster Distributions Through Kernel Density Estimate to Driving Behaviour Scores: A Complete Data Science Pipeline

The study of driving behaviour is an element of great interest for all those companies that provide a fleet management service. Careful observations allow us to find imperfections that, if corrected, entail a reduction of risks and an increase in profits. In order to objectively evaluate this type of behaviour, there are countless indicators to be taken into consideration. In this work, the most significant indicators among those available are used to produce an evaluation of the driving behaviour through a pipeline of operations related to the world of data science. ?? This thesis focuses on the last stage of the process. Machine learning techniques are used to explore a cluster distribution resulting from previous stages. Then that distribution is exploited to calculate a final Key Performance Indicator (KPI) for each trip to be evaluated. This is not done through conventional techniques, such as comparison with a reference approved by a domain expert, but through the application of a Kernel Density Estimation (KDE) function that allows us to produce an assessment based on proximity to the majority of similar behaviours. ?? The results of the experiments demonstrated the validity of the automatically selected set to be considered as a reference, thus confirming the reliability of the KPIs. Furthermore, the analyses on the various types of features have highlighted which are the indicators that most influence the evaluation of driving behaviour.