Performance Comparison of Grid Search and Random Search Methods for Hyperparameter Tuning in Extreme Gradient Boosting Algorithm to Predict Chronic Kidney Failure

Abstract

The kidneys have an essential role in the body; if it does not work properly, it will cause disease, one of which is chronic kidney failure (CRF). Therefore, a machine learning algorithm is needed to help predict CRF, one of which is the Extreme Gradient Boosting (XGBoost) algorithm. However, XGBoost has one thing that must be considered, namely the presence of hyperparameters that need to be tuned. In this research, tuning was done using grid search and random search methods, which would be compared to get a good performance and help predict GGK. In the dataset used, several features were incomplete, so that data cleaning was carried out. Based on the unbalanced amount of data, oversampling with SMOTE and random oversampling methods were needed to balance the data. Next, the data were normalized using two: min-max and z-score normalization. After the data had been normalized, a splitting process was performed with a division of 70% for training data and 30% for testing data. Furthermore, the hyperparameter was tuned to obtain the optimal parameter value, which was then performed data processing and evaluated to obtain the accuracy and f-measure values of the tuned parameter values. In the grid search method with data performed with min-max and z-score normalization combined with oversampling random oversampling, the best results were obtained with an accuracy value of 99.28% and an f-measure value of 0.9942. Comparison with other methods was carried out and the proposed method obtained an accuracy of 99.33% and f-measure 1.0.