Abstract: Graph-based data present unique challenges and opportunities for machine learning (ML). Traditional message-passing Graph Neural Networks (GNNs), while effective in capturing graph topology, face challenges such as limited capture of long-range dependencies and the problem of over-smoothing, where node representations can become indistinguishable after multiple layers of processing. The focus of GNN research has been on maximizing accuracy, and we contend that, as a result, they are not adaptable to a wide range of applications. GNNs struggle in resource-constrained environments like IoT devices, where factors such as limited power capacity, restricted memory availability, constrained CPU resources, data availability, and latency sensitivity are critical. They also are not flexible enough to accommodate power-accuracy trade-offs required for the Green AI initiative that promotes environmentally sustainable AI practices.
Addressing this, we introduce Directed Virtual Coordinates (DVC) and
Topological Coordinates (TC), optimized for low-resource settings like
IoT devices. Our evaluation on OGBN-products and proteins datasets
shows that DVC and TC achieve performance comparable to traditional
models but with orders of magnitude fewer parameters.
This not only aligns with the Green AI initiative, but
also meets the practical need for deploying large and complex
graph analysis algorithms in resource-limited scenarios.
In this presentation, I will discuss our work on proposing these graph
embedding methods for static graphs and explore future directions in
dynamic graph applications.

Adviser: Anura Jayasumana
Co-Adviser: N/A
Non-ECE Member: Michael Kirby
Member 3: Edwin Chong
Addional Members: Indrakshi Ray, Randy Paffenroth

Publications:
Z. Qin, R. Paffenroth and A. P. Jayasumana, "Graph Coordinates and Conventional Neural Networks - An Alternative for Graph Neural Networks," IEEE BigData International Workshop on Data-driven Science for Graphs: Algorithms, Architectures, and Application, Sorrento, Italy, Dec. 2023.

Program of Study:
CS 533
ECE 514
ECE 520
MATH 569
ECE 658
GSTR 600
ECE 799
N/A