Do pollutant interaction terms improve water quality index classification? An empirical study using support vector machine and random forest on Malaysian river data

Abstract

This study examines whether engineered interaction effects among the six main parameters of the Malaysian Water Quality Index (WQI) can improve multi-class water quality classification. River monitoring data from the Department of Environment from 2020 to 2022 were pre-processed and statistically validated prior to model development. Six interaction terms were constructed and evaluated using Support Vector Machine (SVM) and Random Forest (RF) classifiers. Model training was performed using repeated 10-fold cross-validation with Cohen’s Kappa as the optimisation criterion, followed by evaluation on the testing set data. Results showed that models built using only the six main WQI parameters consistently outperformed those incorporating interaction terms. The SVM main effects model achieved the highest performance, with 95.02 per cent accuracy and a Kappa value of 0.9105, indicating an excellent multi-class reliability. Interaction-based models produced a lower Kappa value, suggesting that both SVM and RF were already able to capture complex non-linear relationships without the need for engineered interaction terms. Class-level metrics further demonstrated strong performance across major WQI classes, while reduced sensitivity in Class IV reflected severe class imbalance. Overall, the findings indicate that explicit interaction effects do not improve WQI prediction and that a parsimonious model using only the six WQI parameters is sufficient for reliable classification.