Pandhre, Supriya and Balasubramanian, Vineeth N
(2018)
Understanding Graph Data Through Deep
Learning Lens.
Masters thesis, Indian Institute of Technology Hyderabad.
Abstract
Deep neural network models have established themselves as an unparalleled force in the domains
of vision, speech and text processing applications in recent years. However, graphs have formed a
significant component of data analytics including applications in Internet of Things, social networks,
pharmaceuticals and bioinformatics. An important characteristic of these deep learning techniques
is their ability to learn the important features which are necessary to excel at a given task, unlike
traditional machine learning algorithms which are dependent on handcrafted features. However,
there have been comparatively fewer e�orts in deep learning to directly work on graph inputs.
Various real-world problems can be easily solved by posing them as a graph analysis problem.
Considering the direct impact of the success of graph analysis on business outcomes, importance of
studying these complex graph data has increased exponentially over the years.
In this thesis, we address three contributions towards understanding graph data: (i) The first
contribution seeks to find anomalies in graphs using graphical models; (ii) The second contribution
uses deep learning with spatio-temporal random walks to learn representations of graph trajectories
(paths) and shows great promise on standard graph datasets; and (iii) The third contribution seeks
to propose a novel deep neural network that implicitly models attention to allow for interpretation
of graph classification.
v
Actions (login required)
 |
View Item |
Altmetric
Altmetric