Clustering by transmission learning from data density to label manifold with statistical diffusion

Abstract

Owing to the tremendous diversity and complexity of data in today’s world, some new insights for clustering on data are often desired by developing an alternative to the existing clustering approaches. In this paper, based on the new concepts of the Bayesian transmission system and its transmission learning, a label manifold-based transmission learning machine for clustering (LMTLMC) is accordingly developed. As the first attempt to explain the clustering behavior in a lifelike way, LMTLMC is well justified by revealing the natural parallel between its gradient-based optimization process and the statistical diffusion in statistical physics through the modified Fick’s diffusion law for clustering. Practically, LMTLMC is distinctive in its easy implementation in terms of its global analytical solution, its easy parameter settings and its stable and efficient clustering results. Extensive experiments on synthetic datasets and real datasets demonstrate the promising performance and superiority of LMTLMC for clustering tasks, in contrast to the existing clustering algorithms.