A globally linear model, as implied by conventional Principal Component Analysis (PCA), may be insufficient to represent multivariate data in many situations. It has been known for some time that a combination of several "local" PCA's can providea suitable approach in such cases [1, 2]. An important question is then how to find an appropriate partitioning of the data space together with a proper choice of the local numbers of principal components (PC's). In this contribution we address bothproblems within a density estimation framework and propose a probabilistic approach which is based on a mixture of subspace-constrained Gaussians. Thereby the number of local PC's depends on a global resolution parameter, which represents the assumednoise level and determines the degree of smoothing imposed by the model. As a consequence the model leads to an automatic resolution-dependent adjustment of the optimal principal subspace dimensionalities, which may vary among the different mixturecomponents. Furthermore it allows to provide the optimization with an annealing scheme, which solves the initialization problem and offers an incremental model refinement procedure. Experimental results on synthetic and high-dimensional real-world dataillustrate the merits of the proposed approach.
展开▼
