Indefinite kernel spectral learning

作者：

Highlights：

• Introducing indefinite kernels in multi-class semi-supervised kernel spectral clustering (MSS-KSC) and kernel spectral clustering (KSC, as a special case of MSS-KSC).

• Feature space interpretations for indefinite KSC and indefinite MSSKSC are given.

• Based on the feature space interpretations, the scalability of the proposed methods is discussed using Nyström approximation.

• The proposed methods are evaluated on both small and large-scale real-life datasets. The results imply that, for some unsupervised and semi-supervised tasks, a proper indefinite kernel can significantly improve the performance from PSD ones.

摘要

•Introducing indefinite kernels in multi-class semi-supervised kernel spectral clustering (MSS-KSC) and kernel spectral clustering (KSC, as a special case of MSS-KSC).•Feature space interpretations for indefinite KSC and indefinite MSSKSC are given.•Based on the feature space interpretations, the scalability of the proposed methods is discussed using Nyström approximation.•The proposed methods are evaluated on both small and large-scale real-life datasets. The results imply that, for some unsupervised and semi-supervised tasks, a proper indefinite kernel can significantly improve the performance from PSD ones.

论文关键词：Semi-supervised learning,Scalable models,Indefinite kernels,Kernel spectral clustering,Low embedding dimension

论文评审过程：Received 21 February 2017, Revised 18 October 2017, Accepted 14 January 2018, Available online 3 February 2018, Version of Record 3 February 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2018.01.014