Computer Science Thesis Proposal
Location:
In Person and Virtual - ET
-
Traffic21 Classroom, Gates Hillman 6501 and Zoom
Speaker:
MIN JI YOON
,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://minjiyoon.xyz
Toward more Practical Deep Learning on Graphs
In recent years, Deep Learning on Graphs (DLG) has broken ground across diverse domains by learning graph representations that successfully capture the underlying inductive bias in graphs. However, these groundbreaking DLG algorithms sometimes face limitations when applied to real-world scenarios.
First, domain expertise and tedious work are required to find optimal algorithms for each application. Some algorithms are specialized to specific graph types (e.g., a set of small-sized graphs like molecule graphs) thus not proper for other graph types (e.g., one large-scale graph like e-commerce networks). Even after we narrow down which DLG algorithm to use, practitioners are then faced with laborious hyperparameter tuning.
Second, a few assumptions made in academic research hamper the direct adaptation of DLG models to real-world problems. For instance, social graphs are many times billion to trillion scaled, while many DLG algorithms still assume full batch training.
Finally, real-world graphs are mostly proprietary, while many DLG algorithms often assume they have full access to external graphs to learn their distributions or extract knowledge to transfer to other graphs. In this thesis, I propose to make DLG more practical across four dimensions: 1)automation, 2) generalization, 3) scalability, and 4) privacy.
First, we automate algorithm search and hyperparameter tuning under the message-passing framework. Then we propose a neural module that enables existing DLG algorithms specialized to molecule graphs to generalize to diverse types of graphs. To handle scalability issues, we propose to sample each node’s neighborhood to regulate the computation cost while filtering out noisy neighbors adaptively for the target task. Finally, we redefine conventional problem definitions, including graph generation and transfer learning, to be aware of the proprietary and privacy-restricted nature of real-world graphs.
Thesis Committee:
Christos Faloutsos (Co-chair)
Ruslan Salakhutdinov (Co-chair)
Tom M. Mitchell
Jure Leskovec (Stanford University)
In Person and Zoom Participation. See announcement.