Meaning in Structures: What lower manifold graph network embeddings tell us

Date
2020
Editor
Journal Title
Volume
Issue
Journal Title
Journal ISSN
Volume Title
Abstract
As the internet becomes more accessible around the world, the different networks around the internet such as p2p, email etc. also keeps growing in size. Therefore, analyzing large networks for the purpose of knowledge discovery, in order to learn different governing patterns that dictate these large networks has become quite necessary. With that goal in mind, in this study we studied nine differently sized networks. These varied from small to large networks such as more active human driven networks like email communications, social networks to more passive networks such as p2p. We employed t-distributed stochastic neighbor embedding (t-SNE) on 12 different Centrality Measurement of the networks to find the low dimensional manifold structures of these networks. Spherical k-means was also used to find patterns in those low dimensional manifold representations. We have also analyzed these patterns through an ablation study. We discovered that the manifold structures are partially conserved and same type networks (p2p network) have similar structures within certain error margin, while networks of different types (social network vs p2p network) will be structurally different. With these results in mind we hypothesize that network structures are more affected by the behavior of the nodes/users of the networks instead of having a predefined shape. Based on our findings, we propose a hierarchical categorization of networks in a broader sense, such as communication networks, have hierarchies within their structures, where we can observe the structures changing in a certain pattern or trend. Finally, we have named this as a galaxy model for communication for its self-repeating pattern in large networks. As per galaxy model assumptions, this kind of behavior based unsupervised knowledge discovery methods can help us find further meaningful patterns in large random human networks, which than can be further used to identify and generalize different networks such as migration networks, criminal networks and or corruption networks.
Description
Keywords
Citation
Department Name
Electrical and Computer Engineering
Publisher
North South University
Printed Thesis
DOI
ISSN
ISBN