Constrained Clustering with Seeds and Term Weighting Scheme

While traditional unsupervised learning is blind and the performance relies on the choice of initial seeds. The idea of constrained clustering can use a small number of labeled instances to partly guide a large number of unlabeled instances. It focuses on a set of predefined classes with an aim is to increase the performance of supervised and unsupervised learning using constraints. This paper proposes a new idea of semi-supervised learning based on particularly seeded constrained clustering, where the clustering guidance comes from the statistics of a small set of labeled data. In contrast with existing approaches in seeded K-Means where the labeled instances are specified. However, the proposed work investigates how weighting obtained from a training set affects the seeded-clustering results. Experimental results are demonstrated on three groups of term-weighting statistics; in-collection, intra-class, and inter-class based on frequencies/distributions and an ambiguity class pass entropy value. Text datasets is studied in our experiment. The result also depicts that the term weighting scheme is a potential mean to control/guide the initial and clustering process over a standard normal term weighting scheme.