Abstract:

An improved algorithm was proposed to overcome the shortcomings of the existing local linear embedding algorithm that was not suitable for the nonuniform distribution data and not use the information of distant points. First, to improv the accuracy of the algorithm the geodesic-distance was introduced into the new algorithm in order to take advantage of the information of distant points, and then the harmonic-mean geodesic-kernel matrix was constructed by using the harmonic-mean standardization,which could process robustly non-uniform distribution data. The results of the experiments on UCI data sets showed that the improved algorithm could obtain better performance than the classical local linear embedding algorithm on dimension reduction.

Key words: local linear embedding; harmonic mean; kernel trick; geodesic; manifold learning