10 years of UK traffic data. Classification task to identify fatality of an accident
This kaggle data set can be found here (https://www.kaggle.com/tsiaras/uk-road-safety-accidents-and-vehicles). It covers 10 years of UK traffic data.
As a binary classification task I wanted to predict if an accident was fatal using given information about roads, weather, time of the day, district etc. I used the following algorithms as base models:
Logistic RegressionRandom Forest ClassificationXG Boost
Using hyperopt (http://hyperopt.github.io/hyperopt/), a bayesian hyper-parameter optimizer, I tunned all models and picked the winner based on the ROC AUC score. I was also looking for the best cut-off value by using the geometric mean (optimal point on the ROC AUC curve).
Spoiler: The XG Boost model is the (current) winner with a ROC AUC score of .78.
Additionally I tried to identify the main drivers behind the observed fatality. Using the winning model, it seems that speed (restriction) is the main factor for traffic fatality.