<p>In this paper, we generalize the entropical optimal transport to the free entropical optimal transport, and use the latter as our tool for studying the nonpolynomial (NP)-hard problem of <i>K</i>-means. In fact, the semi-free entropical optimal transport conducts us to introduce a new entropical cost function of <i>K</i>-means for the clustering of a dataset or more generally for finding some good approximation of <i>K</i>-barycenters of a general probability distribution, instead of the least square cost in the classic <i>K</i>-means. The new entropical cost function leads naturally to a new probabilistic algorithm (called entropical <i>K</i>-means or <i>EK</i>-means simply, a probabilistic regularization of Lloyd’s <i>K</i>-means), which is different from the known ones, and very fast in computation. Our free entropical optimal transport approach and algorithm for <i>K</i>-means work not only for distributions of data in the continuum space ℝ<sup><i>d</i></sup>, but also for distributions of discrete data.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Entropical K-means for clustering and an alternating algorithm

  • Jieli Ding,
  • Liming Wu

摘要

In this paper, we generalize the entropical optimal transport to the free entropical optimal transport, and use the latter as our tool for studying the nonpolynomial (NP)-hard problem of K-means. In fact, the semi-free entropical optimal transport conducts us to introduce a new entropical cost function of K-means for the clustering of a dataset or more generally for finding some good approximation of K-barycenters of a general probability distribution, instead of the least square cost in the classic K-means. The new entropical cost function leads naturally to a new probabilistic algorithm (called entropical K-means or EK-means simply, a probabilistic regularization of Lloyd’s K-means), which is different from the known ones, and very fast in computation. Our free entropical optimal transport approach and algorithm for K-means work not only for distributions of data in the continuum space ℝd, but also for distributions of discrete data.