Cluster detection and characterisation based on unsupervised fuzzy learning

This paper introduces a new fuzzy procedure for clustering unlabeled data sets of the form X ={x1,x2,...,xn}⊂ ℜp, called cluster detection and characterisation (CDC). This procedure consists of two main steps followed by a validation module. Using a similarity measure and an associated threshold, th, the first step rapidly explores the object vectors of X in order to find the number c of clusters they form. It provides, in addition to c, a good yet not necessarily optimal prototype for each detected cluster. The second step is an optimization procedure aimed at improving the so learned prototypes. The validation module is useful when varying th leads to more than one plausible solution. Based on five different validity criteria, this module assesses the quality of each solution, which finally permits to retain only the best one. The effectiveness of this algorithm is demonstrated using both synthetic and real test data sets.