Outlier detection from multidimensional space using multilayer perceptron, RBF networks and pattern clustering techniques

As we know that Outlier detection is one of the important aspects of Data Mining, which generally aims to identify potential outliers from datasets. Outliers may sometimes plays important role while taking effective business decisions. This work provides a study of various outlier detection techniques and compares their effectiveness in terms of number of outlier detection, kappa statistic and mean absolute error. Seven algorithms of different categories were tested on three real world datasets to validate the study. We have used pattern based detection of outliers using Multilayer Perceptron, Radial Basis Function Networks, Naïve Bayes Classifiers and Pattern Clustering techniques viz. K-Means, EM and the Agglomerative Hierarchical Clustering. Experimental results show that the Hierarchical Clustering outperforms all other algorithms in terms of number of outlier detection, whereas Multilayer Perceptron and J48 Decision Tree have the highest Kappa Statistic measure. Performance of EM clustering was worst amongst all the algorithms because it was unable to classify all the instances; whereas the performance of RBF Networks and Naïve Bayes Classifiers was almost same and not so satisfactory in terms of outlier detection percentage, Kappa Statistic and Mean Absolute Error.