Strongly robust feature-voting higher dimensional fuzzy clustering

Supervised learning systems learn well on input-output vector pairs with low noise, but often the output vectors are not available so a clustering form of self-organizing learning on the input feature vectors is required. But noise on the features can cause the distance similarity measure to misclassify. To solve this problem we let the features vote, which tolerates unbounded noise on a minority of features. But the classes may not be linearly separable, as is the case for the famous iris data, so the usual clustering is inaccurate. We embed the feature vectors nonlinearly in higher dimensions, an idea from support vector machines. We next apply a feature separability criterion to eliminate extra features that merely add extraneous noise. Then we apply our method of weighted fuzzy expected value clustering in the higher dimensions. The significance is that we often can accomplish self-organizing learning on noisy data in difficult and linearly nonseparable cases. We use the algorithm on a simple test case and then on the iris data.