Histogram Methods for Unsupervised Clustering

A new algorithm is presented to find clusters in a dataset of points in Rn with no prior knowledge of possible clustering. The algorithm detects clusters in a top down fashion by testing modality of density functions generated from the dataset and splitting the set accordingly. Results on synthetic and text datasets demonstrate that the method is comparable to other established unsupervised learning algorithms, which do in fact require the number of clusters ab initio. The method proves to be particularly suitable for certain distributions and offers a valid alternative in situations where most of the well-known algorithms do not produce consistent results.