Project: Kaggle - BNP Paribas Cardiff Claims Management
Objective: This project attempts to automate personal insurance claim approval mechanism by means of using data intelligence. Given a set of completely anonymous attributes we were to model the probability of a claim to be safely approved without needing any further manual scrutiny. We made extensive use of Bayesian/ Likelihood encoding of different order feature interactions to improve the predictive power of our model. We ended up adopting a three-layer ensemble architechtural design for our final model.
Time Frame: Feb, 2016 - Apr, 2016
Evaluation Metric: Binary Crossentropy/ logarithmic loss
Team: Bishwarup Bhattacharjee - Daniel FG - Jeremy Walthers
Total Participating Teams: 2947
Final Standing: 2nd
Minimum Logloss achieved: 0.42079
Models used:
Xgboost (R/ Python)
NN (Keras)
Random Forest (R/ Sklearn)
Extreemly Randomized Trees (R/ Sklearn)
Support Vector Machines (R)
Vowpal Wabbit
Factorization Machine (C++)
Regularized Greedy Forest (C++)
h2o GBM/RF