Outlier Handling in Clustering: A Comparative Experiment of K-Means, Robust Trimmed K-Means, and K-Means Least Trimmed Squared

Abstract

The presence of outliers in data often disrupts modeling results, especially in population clustering and behavioral analysis. Although there are various clustering algorithms that are robust to outliers, such as DBSCAN and t-SNE, K-Means still has challenges in dealing with them. This paper introduces an optimized K-Means LTS algorithm, which incorporates Least Trimmed Square technique to reduce outliers after clustering process. The outlier trimming process occurs after the clustering process, allowing for trimming within each cluster. This algorithm will be compared with K-Means and Robust Trimmed K-Means (RTKM), which both use outlier pruning. The comparison of these three algorithms will consider performance metrics using Silhouette Score and Davies-Bouldin Index, also the run time processes. As a result, K-Means LTS is consistently shown to perform better than K-Means and RTKM when implemented on ten various datasets. In the future, there may be further developments related to determining the best percentage for trimming outliers.