Semi-Supervised Dimensionality Reduction Method Utilizing Pairwise Constraints and Integrating Similarity and Dissimilarity among Data

Abstract Data preprocessing stages, including data reconstruction and dimensionality reduction, play a pivotal role in influencing subsequent machine learning and data mining endeavors. While supervised dimensionality reduction typically surpasses unsupervised methods through direct label exploitation, the scarcity or high acquisition cost of labeled data in practical scenarios has fueled interest in semi-supervised alternatives. These methodologies ingeniously integrate sparse labeled and abundant unlabeled data, to aid data representation and dimensionality reduction.This study introduces a novel nonlinear semi-supervised dimensionality reduction algorithm, which employs pairwise constraints (must-link and cannot-link) and radial basis functions to obtain similarity and dissimilarity among data points. Leveraging spectral decomposition, our algorithm aims to reconstruct the initial data, thereby exposing intricate nonlinear patterns obscured within high-dimensional datasets. Compared to traditional semi-supervised algorithms, a series of experiments based on real data have verified the effectiveness of the proposed method.