Loading...

Structure Learning From Distributed Noisy Data

Karamzadeh Motlagh, Armin | 2019

831 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 52610 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Motahari, Abolfazl; Manzuri Shalmani, Mohammad Taghi
  7. Abstract:
  8. Probabilistic graphical models have great applications in studying and analyzing realworld data. For instance, these models have been used in reconstructing gene regularity networks. Specifically, learning the edges’ structure of graphical models is of great importance.Knowledge about the underlying structure of a graphical model brings about a valuable framework for the decomposition of the model’s distribution and reveals important information such as dependency among dimensions of samples, etc. Most existing methods for structure learning obtain the underlying structure of the model in a centralized fashion and without considering noise in data. In many applications, data exist in a distributed manner, meaning that they are available in a disjoint set of sources. For big data sets, the cost of transmitting data to a central machine for structure learning in a centralized fashion is high, so it is not practical. Also, in many problems, available data are noisy, and accessing information without noise is not possible. Thus, the need for a method for structure learning from distributed noisy data is of high importance. The goal of this thesis is presenting an efficient method for structure learning of Gaussian graphical models with noisy data in a situation that data are split among independent sources through dimension, and each source has a bandwidth-limited communication link with the central machine. Assuming a Gaussian channel between each source and the central node, we have proposed a novel method for structure learning from samples with no coding scheme in two ends of the link; moreover, we have obtained sample complexity bound of the algorithm
  9. Keywords:
  10. Gaussian Graphical Model ; Structural Learning ; Distributed Data ; Noisy Data

 Digital Object List

 Bookmark