Comparação de tempo e resultado de algoritmos de agrupamento utilizando diferentes distâncias e bases de dados

Abstract

This work is a research analysis of unsupervised machine learning algorithms DBSCAN, K-Means and Mean-Shift using Canberra, Chebyshev, Euclidean, Minkowski and Rodrigues distances with some databases from the UCI repository. It aims to compare the average time that each algorithm takes to process eleven databases with ten different distances. In addition, compare the clusters obtained by DBSCAN and Mean-Shift in relation to K-Means. The results show that when using the number and positions of the clusters obtained in DBSCAN as inputs for K-Means and those obtained in Mean-Shift as inputs for K-means, derive from different clusters but with some equality. However, the equality is higher with the combination of Mean-Shift and K-means results. The lowest time averages were obtained by the K-Means algorithm and the highest by the Mean-Shift algorithm. And overall, the Chebyshev distance was responsible for the lowest time averages in 3 of the 4 methods. For these results, it is necessary to choose the appropriate input parameters to generate a considerable number of clusters in the DBSCAN and Mean-Shift algorithms.