Constrained Intelligent K-Means : Improving Results with Limited Previous Knowledge

Abstract

It is here presented a new method for clustering that uses very limited amount of labeled data, employees two pairwise rules, namely must link and cannot link and a single wise one, cannot cluster. It is demonstrated that the incorporation of these rules in the intelligent k-means algorithm may increase the accuracy of results, this is proven with experiments where the real number of clusters in the data is unknown to the method