Text clustering on computing publication using natural logarithm-based TF-IDF in K-means

Abstract

Text clustering is essential for organizing large textual datasets, yet the impact of term frequency inverse document frequency (TF-IDF) logarithmic base selection remains underexplored. This study compares natural logarithm and base-10 logarithm in TF-IDF weighting applied to K-Means and hierarchical agglomerative clustering (HAC) of scientific publications from Universitas Negeri Jakarta across Information Systems, Computer Science, and Informatics Education departments. Clustering performance was evaluated using the Davies-Bouldin index (DBI), silhouette coefficient, and Calinski-Harabasz Index across k=2-5. Results show that both bases produce structurally equivalent clustering outcomes, with k=2 consistently optimal across all departments based on DBI and Silhouette, and consistent agreement between K-Means and HAC on this configuration. Stability analysis across five independent random seeds confirms zero variance at k=2 for two departments and the highest Silhouette score for the third. However, natural logarithm yields higher term weight magnitudes with wider contrast between dominant terms, enhancing qualitative interpretability of cluster themes. Thus, logarithm base selection does not affect clustering structure but meaningfully influences term weight representation, with natural logarithm preferred for thematic analysis. This provides clear, empirically grounded guidance for TF-IDF-based text clustering pipelines. This approach can be further extended to hybrid weighting techniques and additional clustering algorithms.