High Dimensional Sparse Learning in Distributed Systems

Please enable javascript in your browser.

Borghi, Hanieh | 2021

483 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 54005 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Motahari, Abolfazl
Abstract:
Distributed learning has been a popular area of machine learning for researchers according to having access to an unprecedented amount of data distributed over many clients in a network. Moreover, high dimensional learning, when the dimensions of data are high but the effective features are low, has great usage in learning especially in medical science. In this paper, we attempted to develop a method for learning high-dimensional sparse problems in a distributed manner, whit a star shape network. Our main focus in this research is on optimizing the communication flow in the network. The prominent idea of our proposed method is to extract the main feature in the first step, then continuing the process of learning by just these dimensions. In addition, we analyze this work and give some probabilistic grantees, which can be compared with the central manner of the problem. Afterward, we validate our method in synthetic data and real data relating to medical learning to corroborate our theoretical results
Keywords:
Distributed Data ; Communication Cost ; Least Absolute Shrinkage and Selection Operator (LASSO) Estimator ; High-Dimensional Sparse Representation ; Main Feature Extraction ; Sparse Model

No TOC