Unraveling the Parameters of K-Means Clustering for Enhanced Textual Analysis

Abstract

Text classification is a crucial operation for the management of voluminous textual data, the sheer quantity of data in different establishments (companies - universities…) have made their classification into homogeneous and meaningful clusters paramount and to perform this task the K-means algorithm remains the most widely used choice for its low resource consumption and remarkably accurate results. In this study we attempt to explore the parameter configurations of the K-means method by analyzing two distinct text corpora: one from sports articles and the other from news articles. Through meticulous experimentation, we investigate the impact of text pre-processing, similarity measures and pre-trained linguistic models in order to specify the best configuration for using K-means in text classification. The results underline the importance of pre-processing, identify cosine similarity as optimal and highlight the potential of pre-trained models. Our insights provide concrete recommendations for practitioners, advancing text clustering methodology for a variety of applications.