Clustering based outlier detection in fuzzy SVM
Rahul K. Sevakula, Nishchal Kumar Verma · 2014
Fuzzy Support Vector Machine (FSVM) has become a handy tool for many classification problems. FSVM provides flexibility of incorporating membership values to individual training samples. Performance of FSVM largely depends on how well these membership values are assigned to the training samples. Recently, a new approach for assigning membership values was proposed, where only possible outliers are allowed to have membership value lower than `l'. For doing the same, first DBSCAN clustering is performed to find the set of possible outliers and such possible outliers were then assigned membership values based on some heuristics. All other remaining samples were assigned a membership value of `l'. This paper extends the same approach by further analyzing the algorithm, introducing Fuzzy C-Means clustering based heuristic for assigning membership values and also comparing two methods of finding optimal parameters for FSVM model. Experiments have been performed over 4 real world datasets for comparing and analyzing the different methods.