The Quality of Clustering Data Containing Outliers

Abstract

This article evaluates the efficiency and performance of both clustering algorithms: an agglomerative hierarchical clustering AHC (with various linkage options and distance measures) and the $K-Means$ algorithm. We assess the quality of clustering using Davies-Bouldin and Dunn cluster validity indices. Our goal is to compare and analyze outlier detection algorithms depending on the applied clustering algorithm. We also wanted to verify whether the quality of clusters without outliers is higher than of those with outliers. In our research, we compare the LOF (Local Outlier Factor) and COF (Connectivity-based Outlier Factor) algorithms for detecting outliers (selecting $1\%$, $5\%$, and 10% of the most outlier instances in a given dataset). Next, we analyze how clustering quality has improved after excluding such outliers. In the experiments, three real datasets were used with a different number of instances. We wanted to investigate whether it is essential what clustering algorithm and outlier detection method we will use? Our goal was to check whether the clustering parameters impact the obtained clustering results. To the best of our knowledge, no research would combine these issues in one study.