Robust Cluster Approach in Handling Extreme Data on Segmentation of Lecturer Publication Productivity Based on SINTA Data

Abstract

Lecturer publication data on the SINTA portal are typically right-skewed and contain extreme values (outliers) that can distort clustering results, particularly in centroid-based algorithms such as K-Means. This study aims to segment lecturer publication productivity while handling outliers using a robust clustering approach. The data consist of 99 lecturers from Royal University, retrieved from the SINTA portal, with annual publication counts from 2020 to 2025 as features. The research stages include data cleaning, Z-Score normalization, exploratory analysis using box plots, outlier handling with the IQR method, determination of the number of clusters using the elbow method and silhouette score, K-Means segmentation, and evaluation using the Davies-Bouldin Index (DBI), silhouette score, and Calinski-Harabasz Index. The results show that the optimal number of clusters is three (k = 3). Outlier handling improved cluster quality, with DBI decreasing from 1.0837 to 0.6636 and the silhouette score increasing from 0.3877 to 0.4032. The Calinski-Harabasz Index decreased slightly from 66.7273 to 61.5589, which is attributed to reduced between-cluster variance after extreme values were removed. Cluster distribution by study program shows that computing-related programs dominate one cluster, while non-computing programs concentrate in another. These findings can support institutional policy, such as targeted publication training for low-productivity groups and mentoring roles for high-productivity groups.