Perbandingan Kinerja K-Medoids dan Improved K-Medoids Berbasis Crow Search Algorithm pada Klasterisasi Data Transaksi Penjualan Berdasarkan Silhouette Score dan Efisiensi Komputasi

Abstract

The development of digital transaction systems generates large amounts of data that need to be processed into meaningful information to support decision-making. One approach that can be used to analyze consumer purchasing patterns is clustering. The K-Medoids algorithm is known for its robustness against outliers; however, its iterative medoid search process leads to relatively high computational time. To address this limitation, an improved K-Medoids based on the Crow Search Algorithm (CSA) is employed, utilizing a metaheuristic optimization mechanism to determine optimal medoids. This study aims to compare the performance of the K-Medoids algorithm and the improved K-Medoids based on the CSA in transaction data clustering in terms of cluster quality and computational efficiency. The dataset used was obtained from a Point of Sale (POS) system of a fast-food restaurant and consisted of 18,814 transaction records. The research stages included data preprocessing, clustering using both methods, and performance evaluation based on the Silhouette Score and computation time. The results showed that both methods produced the same optimal number of clusters, namely K = 4. The K-Medoids algorithm achieved the highest Silhouette Score of 0.557791, while the improved K-Medoids based on the CSA obtained a Silhouette Score of 0.537240. In terms of efficiency, the improved K-Medoids based on the CSA required significantly shorter and more stable computation times than the conventional K-Medoids algorithm. These findings indicate a trade-off between clustering quality and computational efficiency, implying that the choice of method can be adjusted according to analytical requirements. The main contribution of this study is providing a comparative analysis of the K-Medoids algorithm and the improved K-Medoids based on the CSA on transaction data by jointly evaluating cluster quality and computational efficiency. The findings provide practical recommendations for selecting clustering methods according to analytical requirements and serve as a reference for future research on transaction data clustering.