Local global projection matrix: A novel dimensionality reduction for data classification

Abstract

Dimensionality reduction is essential for preparing high-dimensional data for classification tasks, and it enhances the dataset by removing redundant features, which boosts the efficiency of classifiers. The success of any dimensionality reduction method relies on its numerical stability, but it suffers from overfitting risks and sensitivity to noise. Thus, this paper introduced the Local Global Projection Matrix with Sequential Convolutional Neural Network (LGPM + SCNN) to reduce the dimensionality of the input variables for data classification using image, data, and signal. Initially, the input image, data, and signal are collected from three different databases, which are passed to the LGPM block that contains two stages known as local kernel matrix construction and global aggregation matrix. Here, the outputs from this local kernel construction are aggregated by the Deep Maxout Network (DMN) in the global aggregation stage. Then, this aggregated output is fed to the data classification phase, in which the Sequential Convolutional Neural Network (SCNN) is utilized to classify the data. The LGPM + SCNN acquired an accuracy of 93.28% for 25% outliers, 92.88% for noise density of 0.5 dB, 93.12% for 25% of sparsity, 94.00% for k-fold 9, and 93.60% at 90% of training data for input data.