We present a general framework of semi-supervised dimensionality reductionfor manifold learning which naturally generalizes existing supervised andunsupervised learning frameworks which apply the spectral decomposition.Algorithms derived under our framework are able to employ both labeled andunlabeled examples and are able to handle complex problems where data formseparate clusters of manifolds. Our framework offers simple views, explainsrelationships among existing frameworks and provides further extensions whichcan improve existing algorithms. Furthermore, a new semi-supervisedkernelization framework called ``KPCA trick'' is proposed to handle non-linearproblems.
展开▼
