MIT Libraries homeMIT Libraries logoDSpace@MIT

MIT
View Item 
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Theses - Dept. of Electrical Engineering and Computer Sciences
  • Electrical Engineering and Computer Sciences - Ph.D. / Sc.D.
  • View Item
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Theses - Dept. of Electrical Engineering and Computer Sciences
  • Electrical Engineering and Computer Sciences - Ph.D. / Sc.D.
  • View Item
JavaScript is disabled for your browser. Some features of this site may not work without it.

Graph guided predictions

Author(s)
Garg, Vikas,Ph. D.(Vikas Kamur)Massachusetts Institute of Technology.
Thumbnail
Download1227518830-MIT.pdf (2.014Mb)
Other Contributors
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science.
Advisor
Tommi S. Jaakkola.
Terms of use
MIT theses may be protected by copyright. Please reuse MIT thesis content according to the MIT Libraries Permissions Policy, which is available through the URL provided. http://dspace.mit.edu/handle/1721.1/7582
Metadata
Show full item record
Abstract
Graphs provide a natural abstraction to model relational and strategic data in domains as diverse as biology (e.g., molecules), multiagent settings (e.g., online vendors on ecommerce platforms), and distributed systems (e.g., Internet). Graphs also find much use as theoretical objects (e.g., probabilistic graphical models), and several important algorithms (e.g., max-flow algorithm for image segmentation) can be used when tasks are formulated in terms of graphs. In this thesis, we focus on three important issues that arise when using graph structured data for prediction: (a) compressing graphs to facilitate predictions, (b) understanding the capability of state-of-the-art algorithms (graph neural networks) operating on graphs, and (c) inferring interaction graphs so as to predict strategic outcomes. Our approach to graph compression builds on the idea of optimal transport specifying the cost of mapping a large graph to a smaller one. The cost decomposes as a flow on the edges, and the selection of the subgraph to retain can be optimized via convex Boolean relaxations. Graph neural networks (GNNs) are naturally suited for making predictions based on graphs but they remain poorly understood in terms of what they can and cannot do. We analyze whether GNNs can distinguish graphs that differ in properties such as cycles, but have similar local structure. We also investigate data dependent generalization bounds for GNNs. In many cases the graph structure is not given but must be inferred. This is the case, for example, in trying to understand how a set of players arrive at their decisions. We study the role of structure -- interaction graph -- in the context of predicting outcomes of such games. We analyze conditions under which players converge to an equilibrium, when the samples from equilibrium are themselves at (near) equilibrium, and when the unknown interaction graph can be identified from a data set of sampled context-dependent outcomes.
Description
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, September, 2020
 
Cataloged from student-submitted PDF of thesis.
 
Includes bibliographical references.
 
Date issued
2020
URI
https://hdl.handle.net/1721.1/129251
Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Publisher
Massachusetts Institute of Technology
Keywords
Electrical Engineering and Computer Science.

Collections
  • Electrical Engineering and Computer Sciences - Ph.D. / Sc.D.
  • Electrical Engineering and Computer Sciences - Ph.D. / Sc.D.

Browse

All of DSpaceCommunities & CollectionsBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

My Account

Login

Statistics

OA StatisticsStatistics by CountryStatistics by Department
MIT Libraries homeMIT Libraries logo

Find us on

Twitter Facebook Instagram YouTube RSS

MIT Libraries navigation

SearchHours & locationsBorrow & requestResearch supportAbout us
PrivacyPermissionsAccessibility
MIT
Massachusetts Institute of Technology
Content created by the MIT Libraries, CC BY-NC unless otherwise noted. Notify us about copyright concerns.