Semi-supervised classification in stratified spaces by considering non-interior points using Laplacian behavior

Abstract

Manifold-based Semi-supervised classifiers have attracted increasing interest in recent years. However, they suffer from over learning of locality and cannot be applied to the point cloud sampled from a stratified space. This problem is resolved in this paper by using this fact that the smoothness assumption must be satisfied with the interior points of the manifolds and may be violated in the non-interior points. This fact is based on the property of graph Laplacian in the ϵ-neighborhood of the intersection points. We first generalize this property to NN graph representing the stratified space and then propose a new algorithm that penalizes the smoothness on the non-interior points of the manifolds by modifying the edge weights of the graph. Compared to some recent multi-manifold semisupervised classifiers, the proposed method does not require neither knowing the dimensions of the manifolds nor large amount of unlabeled points to estimate the underling manifolds and does not assume similar properties for neighbors of all data points. Some experiments have been conducted in order to show that it improves the classification accuracy on a number of artificial and real benchmark data sets.

Manifold-based sparse representation for opinion mining

“…This paper focuses on document-level sentiment analysis and attempts to resolve the above challenge by assuming that input sentiment vectors lie on some manifolds. The motivation of this assumption is the line of research that demonstrate high dimensional data lies on some interesting manifolds against to only one manifold 16 , 19 . It means that some near opinions, based on common distance measures, may have various attitudes, and it is needed to another prior knowledge is imposed to resolve the problem.…”

Manifold-based Sparse Representation for Opinion Mining

“…Some methods are linear which can't reflect the non-linear properites of high dimensional sentiment data. There is plenty of research confirming that high dimensional data lie close to the manifold structures [10,11,12,13,14]. Recently, classical linear methods for feature extraction and reduction have been generalized as the nonlinear manifold techniques by establishing a correspondence between a high dimensional space and an intrinsic structure by considering topological relationships.…”