Comparison of Nearest Neighbor (ibk), Regression by Discretization and Isotonic Regression Classification Algorithms for Precipitation Classes Prediction

Selection of classifier for use in prediction is a challenge. In order to select the best classifier comparisons can be made on various aspects of the classifiers. The key objective of this paper was to compare performance of nearest neighbor (ibk), regression by discretization and isotonic regression classifiers for predicting predefined precipitation classes over Voi, Kenya. We sought to train, test and evaluate the performance of nearest neighbor (ibk), regression by discretization and isotonic regression classification algorithms in predicting precipitation classes. A period of 1979 to 2008 daily KMD historical dataset on minimum/maximum temperatures and precipitations for Voi KMD station was obtained. Knowledge discovery and data mining (KDD) process steps were applied. A preprocessing module was designed to produce training and testing sets of files for use with the classifiers. Three classifiers (Isotonic Regression, K-nearest neighbours classifier, and RegressionByDiscretization) were used for training training and testing of the data sets. On running the classifiers the error of the predicted values, root relative squared error and the time taken to train/build each classifier model were computed. Each classifier provided predicted output classes 12 months in advance. Performance of the three classifiers was compared in terms of error of the predicted values, root relative squared error and the time taken to train/build each classifier model. The predicted output classes were also compared to actual classes. Percentage performance for each classifier to actual precipitation classes was computed and compared. The study evaluation showed that the nearest neighbor classifier is a suitable tool for training rainfall data for precipitation classes prediction.