| This thesis proposes and evaluates methods to improve two algorithmic approaches for Hierarchical Agglomerative Clustering. These new methods increase the scalability and speed of the traditional Hierarchical Agglomerative Clustering algorithm without using any approximations. The first method exploits the characteristics of modern Non-Uniform Memory Access architectures, resulting in a parallel algorithm for the stored matrix version of Hierarchical Agglomerative Clustering. The second method uses a data structure called the Cover Tree to speed up the stored data version of the Hierarchical Agglomerative Clustering. For the second method, the thesis proposes both sequential and parallel algorithms. All methods were experimentally evaluated and compared against the state-of-the-art approaches for high performance clustering. The results demonstrate the superiority of the parallel approaches with respect to all baselines and previous work, and the comparison between the stored matrix and the stored data approaches illustrate interesting performance trade-offs. |
