Learning an Embedding for Vehicle Telematics

Abstract

Vehicular telematics involves the collection and processing of data about driving behavior; however, mining and modeling this data is difficult due to its large volume. We hypothesize that the data will follow regular patterns of events that occur during drives, and that we can learn these patterns. With this intuition, we design a neural network that will effectively embed sections of accelerometer data into a lower-dimensional space, with a low loss of information and accuracy of the embedding relative to the dimensionality reduction, as well as several other desirable geometric properties for indexing and analysis of the data. We further develop an accurate summary of the distribution of each drive in this lower-dimensional space, which would serve as a proxy for the occurrence of events within these drives. From this system, we develop a method of comparison between different drives that highlights whether or not particular events occurred in each drive. This could be used to develop a more robust and nuanced risk model, and determine which events in a drive are associated with risk, to provide feedback to end users on their driving.