Evaluation of Partitioning Around Medoids Algorithm with Various Distances on Microarray Data

Data mining is a process which discovers patterns and retrieval knowledge in large datasets. Many learning and data mining algorithms rely on distance metrics. Cluster analysis is one of learning algorithms which adopted to biological data, for example, Microarray expression data. In this study, we assessed the validity of five distance metrics (Euclidean, Manhattan, Minkowski, Cosine and Mahalanobis) with the partitioning around medoids (PAM) algorithm on microarray datasets. Microarray datasets were pre-processed prior to analysis, and the evaluation of the algorithm was undertaken using Krzanowski-Lai validity index. Our results showed When selected microarray datasets were clustered with partitioning around medoids based on Manhattan distance, Minkowski, Cosine and Euclidean distance for different k partitions all exhibited unsatisfactory performance, however, the partitioning around medoids algorithm generates an optimal cluster solution when used with Mahalanobis distance.