Affiliations: [a] State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China | [b] Department of Computer Science and Engineering, University of California, San Diego, La Jolla, CA, USA | [c] Department of Mathematics, The Pennsylvania State University, Pennsylvania, PA, USA | [d] Graduate University, Chinese Academy of Sciences, Beijing, China | [x] Japan Advanced Institute of Science and Technology, Ishikawa, Japan | [y] Nanjing University, Nanjing, China | [z] Osaka University and AFOSR/AOARD, Osaka, Japan
Correspondence:
[*]
Corresponding author: Jun Sun, State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences; P.O. Box 8718, 4# South Fourth Street, Zhong Guan Cun, Beijing 100190, China.
Abstract: We propose a clustering algorithm that effectively utilizes feature order preferences, which have the form that feature s is more important than feature t. Our clustering formulation aims to incorporate feature order preferences into prototype-based clustering. The derived algorithm automatically learns distortion measures parameterized by feature weights which will respect the feature order preferences as much as possible. Our method allows the use of a broad range of distortion measures such as Bregman divergences. Moreover, even when generalized entropy is used in the regularization term, the subproblem of learning the feature weights is still a convex programming problem. Empirical results on some datasets demonstrate the effectiveness and potential of our method.
