Learning the Domain of Sparse Matrices


Salm S., MANGUOĞLU M. , Aktulga H. M.

15th IEEE International Conference on Machine Learning and Applications (ICMLA), California, United States Of America, 18 - 20 December 2016, pp.800-805 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/icmla.2016.124
  • City: California
  • Country: United States Of America
  • Page Numbers: pp.800-805

Abstract

Large sparse linear system of equations arise in many areas of science and engineering. Although, there are several black-box general sparse solvers, usually they are not as effective as domain specific solvers. In addition, most solvers contain multiple choices during the solution process which can be tailored to a specific domain. A natural first step towards a black-box solver that is as effective as domain specific solvers is to come up with a technique to identify the application domain of the problem. In this work, we propose to use some computationally inexpensive matrix properties for the classification task, and apply several classifiers to identify the application domain. Experiments on a large set of sparse matrices show that the domain information is predicted with 75.9% overall accuracy, and matrices in a specific domain can be predicted with 99% accuracy.