Framework for Enhancing the Performance of Classification by RCOS and HiForest

Abstract

This framework includes two novel approaches to choose the outlier from various datasets. First one being Relative Cosine-based Outlier Score (RCOS).It's proposed to measure the deviation score of the objects in which each single attribute deviation is calculated and multiplied to get the entire object deviation. Initially we set the threshold. If the calculated score is greater than the threshold, then the instance is considered as an outlier. These are identified and removed since outliers are not required for classification. Now, the remaining normal objects are subjected to different methods of classification. The second method is Hybrid Isolation Forest (HiForest). It is an enhanced version of isolation forest. Similar to method outliers are identified and removed. An experimental analysis is performed on synthetic real time data sets considered from weka and UCI repository. Classification models are built and the generated results are tabulated and accuracy is recorded. The results obtained by the above methods are compared and graphs are plotted for visualization.