The notebook 1_EDA_FeatureSelection.ipynb contain data cleaning code and a simple exploratory data analysis.
In notebook 2_Classification.ipynb the data has been classified using suggested columns by mutual information ranking, however overfitting of the algorithm is observable.
3rd notebook 3_FeaturesSelectionDropApproach.ipynb contains different approach based on previously carried out experiments, that is selection of numerical features along with dropping of nulls/nans.
By doing so, and disregarding balancing the train dataset with SMOTE/SMOTETomek, XGBoost (w/o hyperparameter tuning) achieved 76% F1 Score, 76% Balanced Accuracy, 87% ROC AUC on a validation set (the set is not highly representative, it contains 200 samples from each group). The results are available in 4_XGB_DropScaleWeights.ipynb.
After hyperparameters tuning, the algorithm performace increased to 80% F1 Score, 80% Balanced Accuracy, and 90% ROC AUC.