Performance Analysis of SMOTE and SMOTEN Techniques for Daily Rainfall Classification using XGBoost

Abstract

The vital function that rainfall patterns fulfill in diverse sectors of life, including agriculture, water management, and disaster mitigation, has engendered the necessity for an accurate rainfall classification system to facilitate early warning and decision-making. However, the development of a classification system is often encumbered by various obstacles, with data imbalance being a prominent one. The objective of this study is to analyze two data resampling techniques, namely SMOTE and SMOTEN, with the aim of improving the performance of the XGBoost classification model. The dataset utilized is accessible on the BMKG website and is classified into five categories. Subsequent to the preprocessing stage, the data is divided by two schemes: 70:30 and 80:20. The determination of the sensitivity of each dataset is achieved through variations in the number of folds in cross validation and the use of learning rates. The experimental results indicate that the SMOTE configuration, with a data division proportion of 80:20 using 10 folds and a learning rate of 0.15, attains the maximum accuracy value of 92.92%. This represents a substantial enhancement from the original dataset accuracy result of 75.36% and surpasses the SMOTE experimental results with an accuracy of 90.58%. Consequently, SMOTEN was found to be superior and effective in managing the imbalance of numerical and categorical datasets, thereby enhancing the performance of the XGBoost model in daily rainfall classification.