A Comparison of K-Means, Hierarchical Clustering, and K-Medoids for Market Segmentation Based on Silhouette Score Evaluation

Abstract

This study compares the performance of K-Means, Agglomerative Hierarchical Clustering, and K-Medoids algorithms for market segmentation using PT XYZ sales data. The dataset consists of the Quantity and Expected Revenue attributes and was processed through data cleaning, currency-to-numeric conversion, invalid data removal, logarithmic transformation, and standardization, resulting in 923 valid records. Clustering was performed using three clusters, representing Low Value, Mid Value, and High Value customer segments. Performance was evaluated using the Silhouette Score, where K-Means achieved 0.4805, Agglomerative Hierarchical Clustering 0.4808, and K-Medoids 0.4840. Although the performance differences among the algorithms were relatively small, K-Medoids achieved the highest score and was therefore selected as the final model. The resulting segmentation consisted of 188 Low Value customers (20.37%), 469 Mid Value customers (50.81%), and 266 High Value customers (28.82%). These findings indicate that K-Medoids provides the best clustering quality while offering greater interpretability through medoid-based cluster centers representing actual data objects. The proposed segmentation can support companies in developing differentiated marketing strategies for low-, medium-, and high-value customer segments.