Optimization of Parametrized Divergences in Fuzzy c-Means

In machine learning the Fuzzy c-Means algorithm (FCM) plays an important role. This prototype based unsupervised clustering method has been extensively studied and applied to a great variety of problems from different research areas like medicine and biology. Commonly the Euclidean distance is used as dissimilarity measure, although any dissimilarity measure would be suited. Recently divergences are used instead [1, 2, 3, 4]. Further, relevance learning, i.e. weighting of input dimensions, was proposed for unsupervised vector quantization to improve cluster separation [5]. We transfer this idea to FCM using generalized divergences for relevance clustering of functional data. We denote this approach as Relevance FCM (R-FCM).