1 September 2011 in Pattern Recognition Letters
3 April 2019 in IEEE Access
The standard formulation of the K-means clustering (Lloyd’s method) performs many unnecessary distance calculations. In this paper, we focus on four approaches that use the triangle inequality to avoid unnecessary distance calculations. These approaches are Drake’s, Elkan’s, Annulus, and Yinyang algorithms. We propose a hybrid MPI/OpenMP parallelization of these algorithms in which the dataset and the corresponding data structures storing bounds on distances are evenly divided among MPI processes. Then, in the assignment step of a K-means iteration, each MPI process computes the assignment of its portion of data using OpenMP threads. In the update step of the iteration, the cluster centroids are computed using a hierarchical all-reduce operation. In the computational experiments, we compared the strong scalability of these four algorithms with the scalability of Lloyd’s algorithm, parallelized using the same approach. The results indicate that all four algorithms maintain an advantage in computing time over Lloyd’s algorithm. A comparison with two software packages, whose sources are publicly available, in the same computing environment, shows that our implementations are more efficient.
1 January 1999 in Lecture Notes in Computer Science
23 June 2010 in Neural Processing Letters
10 January 2015 in Pattern Analysis and Applications
The paper proposes a new method for initialization of the multiple restart EM algorithm for Gaussian mixture model-based clustering. The method initializes randomly both the mean vector and covariance matrix of a mixture component. In particular, the mean vector is initialized by a feature vector selected deterministically from a random subset of candidate feature vectors. The selection criterion is the maximum Mahalanobis distance from the already initialized mixture component centers. The covariance matrix of a component is initialized by randomly generating its eigenvalues and eigenvectors. In computational experiments, the used approach was compared with three other random EM initialization methods. The experiments were performed on synthetic datasets generated from the Gaussian mixtures with the different overlap characteristics, as well as on four real-life datasets. The results on synthetic data indicate that, for well separated clusters, for which the maximum pairwise overlap is not excessively high, the described method yields clusterings which correspond better to the original partitions of data, as indicated by the adjusted Rand index. The experiments on real data indicate that the performance of the method is comparable to other three methods for two smaller datasets and significantly better for two larger datasets.