Application of Bayesian Automated Hyperparameter Tuning on Classifiers Predicting Customer Retention in Banking Industry

Abstract

The paper aims to demonstrate the comparison of accuracy metrics achieved on nine different fundamental Machine Learning (ML) classifiers. Bayesian Automated Hyperparameter Tuning, with Tree-structured Parzen Estimator, has been performed on all of nine ML classifiers predicting the customers likely to be retained by the bank. After visualizing the nature of dataset and its constraints of class imbalance and limited training examples, Feature Engineering has been performed to compensate for the constraints. The ML techniques comprise first using six classifiers (namely, K-Nearest Neighbors, Naive Bayes, Decision Tree, Random Forest, SVM, and Artificial Neural Network––ANN) individually on the dataset with their default hyperparameters with and without Feature Engineering. Second, three boosting classifiers (namely, AdaBoost, XGBoost, and GradientBoost) were used without changing their default hyperparameters. Thirdly, on each classifier, Bayesian Automated Hyperparameter tuning (AHT) with Tree-structured Parzen Estimator was performed to optimize the hyperparameters to obtain the best results on the training data. Next, AHT was performed on the three boosting classifiers as well. The cross-validation mean training accuracy achieved is comparatively quite better than those achieved on this dataset so far on Kaggle and other research papers. Besides, such an extensive comparison of nine classifiers after Bayesian AHT on Banking Industry dataset has never been made before.