Fetch the Kaggle competition data from the Home Credit Default Risk Competition, generate numeric and categorical features then build models using Tensorflow, Scikit-Learn and XGBoost.
The competition dataset contains 8 raw data files. The files include a "loan application" file that information collected at the time of the application. The other 7 files contain supplementary data collect about the loan application from 3rd parties.
For each file in the dataset, we generate min, max, median, mean for each numeric feature. This resulted in 213 features.
For each file in the dataset, we convert all the categorical features into one-hot encoded features. There is a 1:N relationship between the loan application data and supplementary datasets. So there can be many supplementary rows per 1 loan application row. We simply sum the categorical {0, 1} values per loan application id to get the count per category. We then join this dataset to the loan application dataset using the application ID.
This resulted in 191 features
Final training set include numeric and categorical count features from all the data files. The final dataset contained
651 features.
We trained a fully-connected dense neural network using Tensorflow, a random forest model using scikit-learn and a XG Boost Model
I didn't perform a parameter sweep to find optimal model parameters. I used common values seen in kernels where the parameters could run on my consumer-grade laptop in under 20 minutes.
- xgboost_kaggle_higgs_exp/main_plus_all/submissions.csv
- AUC Score 0.68446 (best competition model scored 0.81471)
- Positive Sample Weighting with DNN
- Parameter Sweep with ML models