Semi-Unsupervised Clustering Using Reinforcement Learning

Clusters defined over a dataset by unsupervised clustering often present groupings which differ from the expected solution. This is primarily the case when some scarce knowledge of the problem exists beforehand that partially identifies desired characteristics of clusters. However conventional clustering algorithms are not defined to expect any supervision from the external world, as they are supposed to be completely unsupervised. As a result they can not benefit or effectively take into account available information about the use or properties of the clusters. In this paper we propose a reinforcement learning approach to address this problem where existing, unmodified unsupervised clustering algorithms are augmented in a way that the available sparse information is utilized to achieve more appropriate clusters. Our model works with any clustering algorithm, but the input to the algorithm, instead of being the original dataset, is a scaled version of the same, where the scaling factors are determined by the reinforcement learning algorithm.