Capstone Driver Safety App Project: A Student Perspective
For our Capstone project, we developed three different analyses: the main piece, a Driver Safety Model, as well as two auxiliary research modules, Driver Identification and Hypermiling.
- Driver Safety Model: We combined data collected by a mobile app along with accident claims data to build a model that would predict the likelihood that a driver would get into an accident. The first step (matching the datasets) required developing a fuzzy matching algorithm that could match to-the-minute data on a car’s location, speed, and acceleration rate with accident claims that were at best accurate to the hour. We then compared predictive performance of several models and landed on Random Forest, which could distinguish driver risk such that High-Risk Drivers were more than 4x more likely to get into accidents than Safe Drivers.
- Driver Identification: We needed to build a model that could recognize a driver’s unique driving style so that the app would know if the user is actually behind the wheel of the car he/she is in. To do so, we had colleagues sign up for the app and manually record every time they drove their car or were in a different vehicle. We then built supervised-learning models, trained off their recordings that could classify if a user was actually driving (based on metrics like acceleration rates, turning speeds and radii, frequency of hard stops). We also build an unsupervised model as an alternative (which had less discriminatory power, but was closer to the business objective).
- Hypermiling: We also conducted a research sprint to determine which driving behaviors were the best for saving gas while driving. We borrowed data from Argonne National Laboratory that measured exhaust output against a variety of driving behaviors (acceleration, braking, coasting, top speed) and built a predictive model based on that data.
Presentation:
We gave our final presentation to the executives at our sponsoring company as well as to Northwestern professors. The presentation was well-received and the company accepted our final deliverables in terms of implementation recommendations. We provided the company with all of our technical files (scripts, code, queries; all commented) and a step-by-step explanation of what everything does and how it runs. We also included and our recommendations for next steps.