Abstract
Hierarchical clustering is a common tool for simplification, exploration, and analysis of datasets in many areas of research. For data originating in flow cytometry, a specific variant of agglomerative clustering based Mahalanobis-average linkage has been shown to produce results better than the common linkages.
However, the high complexity of computing the distance limits the applicability of the algorithm to datasets obtained from current equipment. We propose an optimized, GPU-accelerated open-source implementation of the Mahalanobis-average hierarchical clustering that improves the algorithm performance by over two orders of magnitude, thus allowing it to scale to the large datasets.
We provide a detailed analysis of the optimizations and collected experimental results that are also portable to other hierarchical clustering algorithms; and demonstrate the use on realistic high-dimensional datasets.