Scalable Matrix Algorithms for Interactive Analytics of Very Large Informatics Graphs
Abstract
The research developed, implemented, and evaluated traditional matrix and graph algorithms at large scale to allow an analyst with domain insight to explore more interactively the properties of large, e.g., consisting of millions or billions of nodes, social and information networks. Depending on the situation, these larger networks may not fit on a single machine. Although we considered traditional matrix and graph algorithms, e.g., regression and low-rank matrix approximation, we took a nontraditional approach: we extended recent work on randomized matrix algorithms in order to implement them in parallel and distributed environments that are appropriate for networks that are so large that they may be difficult to store on a single machine. By using several complementary methods, this will allow the downstream analyst to test hypotheses as to what properties of the output are realistic properties of the data, given what he or she knows about the sociological mechanisms that generated the data, and what are artifacts of the formalization and algorithms used to explore the data. Currently, this is common for small network analysis, and a result of this research is to make this type of interactive analytics easier for much larger networks.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 14, 2017
- Accession Number
- AD1050770
Entities
People
- Michael Mahoney
- Michael Saunders
Organizations
- Stanford University