A Novel Algorithm to Multi-manifolds Data Sets Classification

The classic manifold learning algorithms are invalid for some data sets which contain multiple non-connected subsets, a new manifolds learning approach is then put forward in this paper. By measuring the connectivity between data points via the minimal connected neighborhood graph, the sub-manifolds are separated correctly. Two key parameters of connecting consumption cost and minimal connected threshold K are used to control the classification procedure. Furthermore, experiments are designed to obtain the experiential parameter formulas of these parameters. The validity of this method is verified by simulation experiment.