Evaluation of a FCM-based FSVM classifier using fuzzy index

Abstract

Support vector machine (SVM), the new machine learning classification algorithm, has shown good generalisation capability in binary classification problems. But, on datasets with outliers or noises, SVM has not shown good classification performance. As fuzzy support vector machine can significantly reduce the effect of outliers or noises, this study has adopted FSVM for model analysis. As membership value on data points influence model performance, fuzzy C-means algorithm was used to evolve membership values on different fuzzy index values. With the new formulation of membership function, new membership values are created and used to run the FSVM model. The computational process of FSVM model on RBF kernel is tested by grid search for different combinations of parameters and the performance of the model on different indices was observed. The classifier with highest classification accuracy for a particular index and kernel parameters is identified as the best classifier for the dataset.