Discrimination Clues From Spectral Clustering

摘要：

Spectral clustering is to partition samples based on spectral decomposition of similarity matrix, so that samples within a same cluster have high similarity and samples from different clusters have low similarity.Linear Discriminant Analysis (LDA) has a similar objective except it performs through reducing the dimensionality of samples. However, when only a small size samples are available for training, the irreversibility of within-class scatter matrix needs to be avoided through select a subspace. The rationality of the selection for almost existed approaches is (empirical)higher recognition rate of selected classifiers. In this paper, we explore how to select a relevant subspace,where the relevance is defined as maximum averaged margin between classes. The relevance is revealed through linking to spectral clustering. The membership of samples from spectral clustering provides us an approach to discrimination between classes. Two typical datasets (i.e., USPS digits and PIE faces) are utilized to visualize the relevance.

关键词： Dimensionality Reduction Discriminant Analysis Feature Selection Face Recognition

作者: Hong TANG Nozha Boujemma Henri Maitre

作者单位: Project IMEDIA, INRIA Paris, France GET/Telecom Paris Paris, France

会议类型: 国际会议

会议名称: The International Colloquium on Onformation Fusion 2007(2007年国际信息融合研讨会)

会议地点: 西安

会议语种:英文

页码: 357-362

在线出版日期: 2007-08-22（万方平台首次上网日期，不代表论文的发表时间）