Porto-Seguro-Safe-Driver-Prediction

Code used for the Kaggle competion: "Porto Seguro's Safe Driver Prediction". This was a predictive analysis project to determine how likely a given driver is to claim insurance in the coming year. Classification algorithms used - Xgboost, Random forest classifier, Artificial Neural Networks and Logistic regression. The train and test data are available on the Kaggle page : https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/data. Training dataset had exterme class imbalance as the number of records with claims is always significantly lower than non-claims. Class balancing approaches used : Undersampling using Kmeans, Random undersampling, Synthetic Minority Oversampling Technique. Feature reduction approaches used : Feature importance obtained from RandomForestclassifier, Logistic regression and Frequency and binary encoding. Best result obtained using the following combination: Class imbalance: Random undersampling. Feature selection: Binary and frequency encoding. Prediction algorithm: Xgboost.