Distributed Structure Learning of Gaussian Graphical Models

Please enable javascript in your browser.

Mirzaeifard, Reza | 2019

614 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 52175 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Manzuri Shalmani, Mohammad-Taghi; Motahari, Abolfazl
Abstract:
Nowadays, the explosion of the volume data provides more accuracy in machine learning models. But, Working with a vast amount of data is not easy, especially in a situation that data are distributed over the systems. In such systems, designing distributed learning algorithms that in communication efficient setting demand reliable and more accurate results, are so important. We studied sparse structure learning of Gaussian graphical model in a situation that our data are distributed over the system and each machine has a dimension of data. Each local machine should send its data to a central machine and the central machine is responsible for learning the structure. For reliable learning under bandwidth limitation, we proposed Signs methods. In Signs method, each machine just sends the signs of data to the central machine. the signs of data are determined according to the mean of the Gaussian model or the middle of the Nonparanormal model. Theoretical results show that with Signs method we can recover the sparse structure of Gaussian graphical model and Nonparanormal model, consistently
Keywords:
Gaussian Graphical Model ; Nonparametric Method ; Structural Learning ; Distributed Learning ; Communicatin Cost

No TOC