Evaluation of optimal clusterings found by cluster validation measures

K-REx Repository

 dc.contributor.author Shi, Yu dc.date.accessioned 2020-08-03T21:35:30Z dc.date.available 2020-08-03T21:35:30Z dc.date.issued 2020-08-01 dc.identifier.uri https://hdl.handle.net/2097/40778 dc.description.abstract There are many measures developed for assessing clustering algorithms. en_US However, little work has been done to determine what type of clusterings these validation measures would consider the best.'' In particular, if a clustering validation measure performs well, then it should be able to identify the correct'' clustering when when presented with all possible ways of clustering a dataset. We evaluate the performance of five clustering validation measures---Silhouette, Hubert-Gamma, R-squared, the Dunn family of indices, and the data Davies-Bouldin index---on five small clustered datasets. To obtain a large set of candidate clusterings, we view each dataset as a graph and form a connected bottleneck subgraph. On this subgraph, we identify all set-connected partitions---those whose blocks are connected---that satisfy a set of constraints on the number of blocks and the size of each block within the partition. We then apply the validation measure on each of the possible partitions to determine the clustering that each validation measure considers to be optimal. Based on test results, we find each measure has its own preferences. For example, the silhouette measure tends to be better at capturing connected regions, and many others measures prefer clusterings that contain many clusters. Finally, we compare the clusterings found by the validation measures to those obtained by other popular clustering methods including $k$-means, hierarchical agglomerative clustering (HAC), density-based spatial clustering of applications with noise (DBSCAN) and ordering points to identify the clustering structure (OPTICS). dc.language.iso en_US en_US dc.subject Clustering en_US dc.subject Cluster validation measures en_US dc.title Evaluation of optimal clusterings found by cluster validation measures en_US dc.type Report en_US dc.description.degree Master of Science en_US dc.description.level Masters en_US dc.description.department Department of Statistics en_US dc.description.advisor Michael J. Higgins en_US dc.date.published 2020 en_US dc.date.graduationmonth August en_US
﻿

Center for the