Drinking water quality data sets used in learning models have been highly imbalanced, which has weakened the prediction ability of models for drinking water quality. Although some efforts have been made to address the issue of imbalance, little is known about the suitable technologies for drinking water quality prediction. Here, a total of 16 common learning models were applied individually to compare the drinking water quality prediction performance based on a large-scale highly imbalanced drinking water quality data set. Our results showed that ensemble, cost-sensitive learning models with higher F1-scores were more suitable for predicting drinking water quality, compared to other models tested in this study. In addition, the learning model performance could be enhanced by the introduction of two mainstream sampling algorithms [synthetic minority oversampling technique (SMOTE) combined with the Tomek links technique (TLTE) or the edited nearest neighbor technique (ENNTE), SMOTE + TLTE or SMOTE + ENNTE, respectively]. In particular, the F1-scores of deep cascade forest (DCF) with SMOTE + TLTE or SMOTE + ENNTE reached 94.54 ± 2.51% and 94.68 ± 2.72%, respectively. As a consequence, DCF with these two sampling algorithms has greater potential to be applied in drinking water quality monitoring and prediction, as well as other fields that have suffered from issues of imbalanced data.

No abstract