Abstract:In this manuscript, we introduce a novel local manifold learning method that extracts further structural information from the high-dimensional data manifold. We construct a neighborhood graph matrix that incorporates two different types of structural information: linear relationships and Euclidean distances between local data points. Based on this neighborhood graph matrix, we propose a new cost function to represent the data manifold in the low-dimensional space. Subsequently, the low-dimensional data manifold is computed by minimizing the proposed cost function with respect to the low-dimensional data points. The similarity of the structure of the low-dimensional represented data manifold to the high-dimensional data manifold is evaluated based on the functionality of the k-nearest neighbor classifier. The proximity of the achieved recognition rate in the high-dimensional space to the calculated recognition rate in the low-dimensional data manifold can serve as a metric of similarity. We demonstrate the effectiveness of the proposed method on various datasets, including the ORL face image database, Parkinson's Disease Classification dataset of UCI machine learning repository, and CinCECGTorso dataset of UCR time series classification archives, as well as three 3-dimensional synthetic datasets. Our results demonstrate that the proposed method surpasses state-of-the-art manifold learning techniques in effectively preserving the data structure. Furthermore, the method displays reduced sensitivity to the initial number of neighbors, contributing to its robustness. Additionally, it effectively mitigates noise effects in specific low-dimensional representation spaces, resulting in a notable improvement in recognition rates by more than 5%.

Learning Bundle Manifold by Double Neighborhood Graphs

Learning Exemplar-Represented Manifolds in Latent Space for Classification.

Locality Pursuit Embedding

Learning an Image Manifold for Retrieval

Deep Manifold Transformation for Nonlinear Dimensionality Reduction

Deep Manifold Computing and Visualization

Deep Manifold Computing and Visualization Using Elastic Locally Isometric Smoothness

Neighbourhood Sensitive Preserving Embedding for Pattern Classification.

Geodesic Based Semi-supervised Multi-manifold Feature Extraction

A Regularized Approach for Geodesic-Based Semisupervised Multimanifold Learning

Active learning on manifolds

Hierarchical Neighbors Embedding

Learning nonlinear manifolds based on mixtures of localized linear manifolds under a self-organizing framework

Neighborhood preserving embedding on Grassmann manifold for image-set analysis

Investigating Manifold Learning Algorithms Based on Magnification Factors and Principal Spread Directions

Geodesic Gabriel Graph Based Supervised Nonlinear Manifold Learning

Learning Manifold Structures With Subspace Segmentations

Manifold learning with local geometry preserving and global affine transformation

DLME: Deep Local-flatness Manifold Embedding

Manifold-based learning and synthesis.

Manifold learning through locally linear reconstruction based on Euclidean distance