Maximum-Linear-Patch based Outlier Detection for Robust Manifold Learning

In this paper, we propose a novel outlier detection method for robust manifold learning. First, we assume that the input high-dimensional data can be represented by an integration of local linear patches, and each patch consists of samples in the local neighborhood that maintains a linear relationship. Then, we use a hierarchical divisive clustering method to seek maximum linear patches (MLPs) and present a local density based scheme to detect outliers in each MLP. In order to evaluate the performance of outlier detection, we also propose an improved outlier detection evaluation method based on manifold distance, which is suitable for robust manifold learning. Last, we give several experiments to demonstrate the effectiveness of the proposed outlier detection method and evaluation method.