Analisis Perbandingan Sensitivitas dan Kinerja K-Means dan DBSCAN Terhadap Variasi Noise

Rafael Austin (Universitas Esa Unggul)
Ricky Dwi Putra (Universitas Esa Unggul)
Ardi Ardi (Universitas Esa Unggul)
Fuzail Fazle Rabbi (Universitas Esa Unggul)
Lucky Pujiono WS (Universitas Esa Unggul)
Vitri Tundjungsari (Universitas Esa Unggul)

TEKNOLOGI INFORMASI
Vol 17 No 1 (2026): JURNAL TEKNOLOGI INFORMASI: Teori, Konsep dan Implementasi

Inaccurate or noisy data presents a significant challenge in machine learning, particularly in unsupervised clustering tasks. This study evaluates the robustness and performance of two popular clustering algorithms, K-Means and DBSCAN, against various levels of Gaussian noise (5%, 10%, 20%, and 30%) injected into a customer dataset. Evaluation was conducted using Silhouette Score and Davies-Bouldin Index (DBI). Initial results indicated that DBSCAN performed slightly better with a Silhouette Score of 0.4817 compared to K-Means at 0.4101. However, after noise injection, K-Means demonstrated superior stability by maintaining more consistent cluster memberships, whereas DBSCAN was more sensitive to distance variations, leading to significant fluctuations in cluster assignments. The study concludes that K-Means is more reliable for datasets where cluster integrity must be preserved despite minor data irregularities.