说明：  Weintroduceanewsimilaritymeasurebetweendatapointssuited respecttoclustercentroids),shareswiththepreviousapproachthe for clustering and classi?cation on smooth manifolds. The pro- sameproblemsin itsoptimization formulation. posed measure is constructed from a dual rooted graph diffusion Recently, the focus of attention in unsupervised learning has over the feature vector space, obtained by growing dual rooted turned to spectral clustering methods due to its many successes minimum spanning trees (MST) between data points. This diffu- [1]. These methods use the spectral content of a similarity ma- sionmodelforpairwiseaf?nitiesnaturallyac