Actionable Clustering

In this paper, we define a new notion for a clustering to be useful, called actionable clustering. This notion is motivated by applications across various domains such as in business, education, public policy and healthcare. We formalize this notion by adding a novel constraint to traditional unsupervised clustering. We argue that this notion is different from semi-supervised clustering, supervised clustering, weighted clustering and cannot be effectively modeled in these frameworks. We elucidate the cases when there is a feasible solution to our problem and propose a new algorithm called actionable kmeans, which is a modification of the standard kmeans algorithm to respect this constraint. We then analyze our algorithm by proving that it converges along with analysis of its time complexity. We empirically portray on synthetic and real data the superior performance of our algorithm relative to this new notion in terms of the value of the unsupervised clustering objective obtained compared with supervised and semi-supervised clustering methods adapted to this setting.