A Multifaceted Approach to Improving the Practicality of Structural Graph Algorithms
Year: 2018
Type: article
Abstract: A Multifaceted Approach to Improving the Practicality of Structural Graph Algorithms. (Under the direction of Blair D. Sullivan.)Graph algorithms have become an integral part of modern data analytics, but existing approaches have struggled to scale to increasing network sizes.The theoretical computer science community has a rich history of research that circumvents these scalability issues through algorithms that exploit the structural sparsity of graphs.Because many real-world networks from diverse domains are known to share properties like sparsity, clustering, and heavy-tailed degree distributions, structure-based algorithms appear on the surface to be an attractive alternative.However, they come with their own set of problems, such as non-constructive proofs, massive constants hidden in big-O notation, and attention paid to exploiting structures that are unlikely to occur in real data.Consequently, there is a large gap between the most theoretically efficient algorithms and the most practical ones.This dissertation focuses on alleviating practical barriers to the use of structural graph algorithms in large-scale data analytics, addressing problems on multiple different fronts.First, we show that some structural features can still be identified even in the presence of missing or unknown data.A second contribution is CONCUSS, a first-of-its-kind implementation of a subgraph isomorphism counting algorithm that exploits the bounded expansion structure of graph classes.Through a thorough experimental evaluation of CONCUSS, we establish conditions under which it is competitive with existing algorithms and highlight ways to enhance the general algorithmic framework.As extensions to this framework, we introduce p-linear colorings-an alternative graph coloring used to identify bounded expansion structure-as well as practical dynamic programming algorithms to perform a variant of local search in fixed-parameter tractable time in graph classes with bounded expansion.Despite these successes, we prove that recognizing classes of bounded expansion by measuring the density of shallow topological minors is likely not possible in polynomial time.Finally, to demonstrate a concrete way in which structure-based algorithms enable better analysis in scientific domains we build a framework for extracting neighborhoods from metagenomic sequencing data in an interdisciplinary collaboration with computational biologists.Overall, this work illustrates multiple effective strategies for bridging the gap between efficient and practical algorithms.
Author: Michael P. O’Brien
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2796655307
