Comparing SMOTE, Cost-Sensitive, and Fuzzy SVM for Day-Ahead Air Quality Category Forecasting in Jakarta
Abstract
This study compares three strategies for handling class imbalance in multiclass support vector machines (SVM) for forecasting the next-day category of the Indonesian Air Pollution Standard Index (ISPU): SMOTE combined with SVM, a cost-sensitive SVM, and a fuzzy SVM for class imbalance learning (FSVM-CIL), with a standard SVM as the baseline. The data are 7,092 daily records from five air quality monitoring stations in Jakarta for 2021–2025, merged with NASA POWER meteorology. Predictors contain only lagged pollutant, meteorological, and calendar information, so no information from the target day enters the model. The Moderate category covers 76.2% of station-days, while the rarest category differs across stations. Models were tuned by time-series cross-validation, tested on 2024 and 2025, and compared with the Friedman test over ten station-year blocks. FSVM-CIL achieved the highest mean macro-F1 (0.521), balanced accuracy (0.601), and G-mean (0.502), against 0.445, 0.458, and 0.175 for the standard SVM, and raised the recall of Unhealthy days from 0.18 to 0.44. The differences were not statistically significant (?² = 5.67, p = 0.129), and the best model varied by station. Imbalance handling improves the detection of rare air quality categories, but the strategy should be chosen by local validation.
Keywords
Share and Cite
Article Metrics

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

