Locally finite distance clustering with discriminative information

Partition-based clustering methods, such as point center-based and plane center-based clustering techniques, have drawn much attention due to their simplicity and effectiveness in general clustering tasks. However, most of these methods use an unbounded distance during clustering, which may cause their performance to be sensitive to the defined infinite measure, and almost all of them cannot automatically identify halos. To solve these problems, by adopting a locally finite capped l2,1-norm distance in clustering, this paper proposes a novel clustering method named locally finite distance clustering with discriminative information (LFDC). The LFDC effectively solves the above problems and realizes robust clustering by solving a series of eigenvalue problems. We test the effectiveness of the LFDC on a number of different data, including artificial data, benchmark data, and image segmentation data. The experimental results show that the LFDC is more robust than the compared methods.