Dimensionality Reduction via Graph-based Feature Selection

Given the rapid influx of data, organizing and managing its features poses a significant challenge. Many features are deemed redundant, impeding efficient storage and training of models. We introduce a method to address this challenge by defining an algorithm that successfully achieves dimensionality reduction. By establishing a threshold, we identify pairs of features with significant linear correlations which are modelled as vertices and edges on a graph. Using graph theory principles, we can pinpoint non-essential features. By strategically removing these points, our method proceeds to streamline data management processes and enhances computational efficiency while outperforming Principal Component Analysis (PCA).