Soft-Voting Classification using Locally Linear Reconstruction

Locally linear reconstruction (LLR) is a crucial step in the dimensionality reduction method called locally linear embedding (LLE), which aims to build a kind of weighted relationships for nearby data points. In this paper, we use this step in a different way to derive a new supervised classifier. The classifier labels a given test sample by checking which class of training samples can best reconstruct that sample. On a set of benchmark data sets, this new classifier performs better than k-nearest neighbor classifier and another state-of-the-art one. And most importantly, the classifier can be used to very large data sets because of the low time complexity.