Abstract
With the availability of large scale computing platforms and instrumentation for data gathering, increased emphasis is being placed on efficient techniques for analyzing large and extremely high-dimensional datasets. In this paper, we present a novel algebraic technique based on a variant of semi-discrete matrix decomposition (SDD), which is capable of compressing large discrete-valued datasets in an error bounded fashion. We show that this process of compression can be thought of as identifying dominant patterns in underlying data. We derive efficient algorithms for computing dominant patterns, quantify their performance analytically as well as experimentally, and identify applications of these algorithms in problems ranging from clustering to vector quantization.We demonstrate the superior characteristics of our algorithm in terms of (i) scalability to extremely high dimensions; (ii) bounded error; and (iii) hierarchical nature, which enables multiresolution analysis. Detailed experimental results are provided to support these claims.