A Computational Comparison of Parallel and Distributed K-median Clustering Algorithms on Large-Scale Image Data

Most commonly used clustering algorithms are those aimed at solving the well-known k-median problem. Their main advantage is that they are simple to implement and use, and they are flexible in choosing dissimilarity measures (not necessarily metrics). K-median algorithms are also known to be more robust to noise and outliers in comparison with k-means algorithms. In spite of that, they have been of limited use for large-scale clustering problems due to their high computational and space complexity. This work aims at computational comparison of k-median clustering algorithms in a specific large-scale setting—clustering large image collections. We implement distributed versions of the most common k-median clustering algorithms and compare them with our parallel heuristic for solving large-scale k-median problem instances. We analyze clustering results with respect to external evaluation measures and run time.