A Unified Framework to Integrate Supervision and Metric Learning into Clustering

In this paper, we propose a unified framework for applying supervision to discriminative clustering, which: (1) explicitly formalizes the problem of training a partition function as a supervised metric-learning process; (2) learns a partition function that can optimize a supervised clustering error; and (3) flexibly learns a distance metric with regard to any clustering algorithm. Moreover, we develop a general gradient-descent learning algorithm that trains a distance metric under this framework. The convergence of this algorithm is guaranteed for some restricted cases. Our experimental study shows the significant performance improvement after integrating supervision and distance metric learning in clustering, trained in our new framework.