Pruning Neighborhood Graph for Geodesic Distance Based Semi-Supervised Classification

Abstract

Recently semi-supervised learning has been gain a surge of interests, but there is a few of research on semi- supervised learning using geodesic distance. The simplest semi-supervised classification algorithm is geodesic nearest neighbors (GNN). However the naive implementation of GNN algorithm is sensitive to the neighborhood scale parameter and suffers from the dilemma of neighborhood scale parameter selection. In this paper, instead of searching for the best neighborhood parameter, we propose a pruned-GNN, which utilize the non-negative reconstructing coefficients to prune the neighborhood graph in order to facilitate the selection of neighborhood scale parameter. Experimental results on several benchmark databases have shown that the proposed pruned-GNN can produce promising accuracies.