Multi Class Unbalanced Text Classification Through Efficient Feature Selection A Machine Learning Approach

Abstract

The exponential growth of digital text on the web has made Text Classification (TC) a critical newlineresearch domain, enabling the automatic organization, retrieval, and analysis of large-scale newlineinformation. Manual labeling of text is not only impractical but also infeasible given the newlinehigh dimensionality, sparsity, and skewed class distributions present in real-world datasets. newlineThe challenge becomes more complex when documents are associated with multiple labels newlineand the class distributions are unbalanced, since conventional models often fail to correctly newlinerepresent minority classes or identify the most relevant features. newlineFeature selection (FS) thus becomes an essential step in TC, aiming to reduce dimensionality, newlineimprove classifier performance, and retain only the most discriminative terms. newlineTo address these challenges, this work introduces a Modified Chi-Square (ModCHI) based newlinefeature selection technique that improves upon the traditional Chi-Square method. The effectiveness newlineof ModCHI was evaluated on the Reuters multi-labeled, multi-class benchmark newlinedataset using four classifiers: Linear SVM (LSVM), Decision Tree (DT), Multilevel KNN newline(MLKNN), and Random Forest (RF). newlineBuilding on this foundation, the ModCHI technique was further integrated with a newlineRandom Vector Functional Link Network (RVFLN) to enhance classification performance. newlineIn addition to generic text classification, this research also explores an important application newlinein the domain of mental health analysis. Mental illness has emerged as a major global health newlinechallenge, significantly reducing productivity and quality of life. In this study, a TOPSISModCHI newlinebased approach was applied with four Randomized Neural Networks (RandNNs) newlinenamely Broad Learning System (BLS), RVFLN, Kernelized RVFLN (KRVFLN), and Extreme newlineLearning Machine (ELM) to detect types of mental illness from social media text (Reddit newlineMental Health Dataset). The findings show that BLS with TOPSIS-ModCHI achieved the newlinebest performance. The results demonstrate the ability of the proposed approach to correctly newlinei