Predicting river water quality index using data mining techniques

Abstract

This paper demonstrates the application of data mining techniques to predict river water quality index. The usefulness of these techniques lies in the automated extraction of novel knowledge from the data to improve decision-making. The popular classification techniques, namely k-nearest neighbor, decision trees, Naive Bayes, artificial neural networks, rule-based and support vector machines were used to develop the predictive environment to classify water quality into understandable terms based on the Overall Index of Pollution. Experimentation was conducted on two types of data sets: synthetic and real. A repeated k-fold cross-validation procedure was followed to design the learning and testing frameworks of the predictive environment. Based on the validation results, it was found that the error rate in defining the true water quality class was 20 and 28%, 11 and 24%, 1 and 38% and 10 and 20% for the k-nearest neighbor, Naive Bayes, artificial neural network and rule-based classifiers for synthetic and real data sets, respectively. The decision tree and support vector machines classifiers were found to be the best predictive models with 0% error rates during automated extraction of the water quality class. This study reveals that data mining techniques have the potential to quickly predict water quality class, provided data given are a true representation of the domain knowledge.