Comparative Study of the State of the Art Machine Learning Algorithms in Predicting Potability of Drinking Water

Water constitutes 60% of the human body; thus, it is a vital part for life. It is facing increasing threats from human activities causing pollution. The depletion of freshwater suitable for consumption necessitates assessing water quality, prompting a study proposing a machine learning analysis of 720 water samples. The evaluation includes physical parameters like turbidity, temperature, and chemical composition such as calcium, sodium, and microorganism presence to predict water potability. Ten machine learning algorithms, including AdaBoost, KNN, random forest, gradient boost, decision trees, Gaussian Naive Bayes, multiple regression, ridge regression, logistic regression, and lasso regression, are employed. The assessment criteria encompass accuracy scores, recall, precision, F1-score, weighted F1-score, mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and R-squared. The study identifies gradient boost as the most accurate model with a 100% accuracy score and minimal errors. Key water parameters for future predictions are highlighted as potassium, magnesium, carbonate, total dissolved salts (TDS), and total hardness (TH). Correlated attributes significantly impact the Gaussian Naive Bayes algorithm. Among regression models, logistic regression excels with an MSE of 0.077 and an R-squared value of 0.633. Tukey's method detects and treats outliers, while min–max scaling optimizes the dataset for model tuning. In conclusion, the study underscores the critical role of advanced machine learning algorithms in predicting water quality, providing valuable insights for future modeling and emphasizing essential parameters for assessment.