Font Size: 

In this paper, we present a new parameter free clustering algorithm for Similarity-based datasets. When the objects are defined by their relations with each other, clustering algorithms must be applied on a matrix of similarities. The proposed approach is based on matrix reordering. We extract a one-dimensional signal of pairwise distances from the reordered matrix and we apply a signal processing method to detect peaks in the signal related to clusters’ separation. In that way, there is no parameter to tweak in order to obtain a reasonable clustering. We compared the approach with state-of-the-art algorithms to evaluate it’s efficiency.