A Robust Locally Linear Embedding Method Based on Feature Space Projection

Abstract

At present, most of the manifold learning methods use the K-Nearest Neighbor (KNN) criterion to determine the adjacency relationship between data points, which is not complete in the description of the sample distribution in the original data space. When many noise data are included in the observation space, the results of constructing adjacency graph by KNN may contain too many areas outside the unsupported domain, which will produce the wrong geometric projection distance. Isometric Feature Mapping (ISOMAP) and Locally Linear Embedding (LLE) algorithm are very sensitive to noise data because of above reason. In view of the above problem, a robust locally linear embedding method based on feature space projection is proposed. Based on LLE algorithm, the feature space projection is introduced to smooth the original samples and to improve the robustness of the noise data set in the process of reducing the dimension using LLE algorithm. The experimental results prove that this method is effectiveness.