Comparative Analysis Of Clustering Algorithms Applicable To Massive Dataset Processing

Cilj ovog rada je ispitivanje i usporadba svojstava algoritama grupiranja prikladnih za grupiranje velikih skupova podataka. Algoritmi odabrani kao prikladni u tu svrhu su: K-means, CLARANS, BIRCH i DENCLUE. Kvaliteta grupiranja je najvažniji kriterij usporedbe algoritama. Kvaliteta grupiranja mjeri se unutarnjim i vanjskim validacijskim metrika. Objašnjen je rad svih odabranih algoritama i validacijske metrike su definirane, a potom su provedena ispitivanja. Algoritmi grupiranja primjenjeni su ponajprije nad sintetiˇckim skupovima podataka, ali i na dva stvarna skupa. Ispitan je utjecaj oblika grupa, neravnoteže med¯u grupama i šuma u podacima na kvalitete grupiranja te skalabilnost algoritama. Algoritmi su primjenjeni i na stvarnim skupovima te je ocijenjena njihova prikladnost.The goal of this paper is testing and comparison of properties of clustering algorithms applicable to massive dataset processing. Algorithms chosen as applicable for this task are: K-means, CLARANS, BIRCH and DENCLUE. Quality of clustering is the most important criterion of comparison between algorithms. Clustering quality is measured by internal and external validation metrics. Each chosen clustering algorithm is explained and each used validation metrics are defined, then testing was conducted. Clustering algorithms were primarily used on syntetic datasets, but also on two real ones. The goal of the conducted tests was to discern sensitivity of clustering algorithms to the shape of clusters, cluster imbalance and clustering of noisy data and their impact on quality of clusterings. Scalability of clustering algorithms was also tested. Algorithms were also used on real datasets and their applicability was discussed