Temporal hypergraph modeling via inter-geometrical learning

Agarwal, Shivam

Permalink

https://hdl.handle.net/2142/124326 Copy

Description

Title

Temporal hypergraph modeling via inter-geometrical learning

Author(s)

Agarwal, Shivam

Issue Date

2024-04-26

Director of Research (if dissertation) or Advisor (if thesis)

• Han, Jiawei
• Peng, Hao

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• graph neural networks
• spatio-temporal learning
• machine learning

Abstract

Hyperbolic geometry has advanced learning representations on graphs with inherently complex geometrical and hierarchical characteristics. However, most real world networks innately comprise of higher-order relations, dynamic behavior, and scale-free temporal characteristics with varying degrees of hyperbolicity. Combating these gaps, we propose THRONE, a temporal, inter-geometrical interaction learning-based hypergraph convolution method to capitalize on the complex, time-varying, higher-order relations and the varying hyperbolicity of network structures. Further, we enhance the hypergraph convolution by applying attention infused with hyperbolic distance information among node and hyperedge representations. THRONE incorporates hyperbolic temporal convolution layers to encode scale-free spatio-temporal information and dynamic time-evolving network structures. We extend THRONE to hypergraph-level tasks by introducing THRONE-Pool, a novel hyperbolic hypergraph pooling method to encode higher-order scale-free networks. Through a series of quantitative and exploratory analyses on ten node-level and six network-level tasks across static, spatio-temporal, and time-evolving dynamic hypergraphs, we demonstrate THRONE's practical applicability in comparison to competitive baselines. We dissect THRONE's performance contributions on a variety of benchmarks and applications spanning finance, health, traffic, wind energy, and citation networks through ablations to highlight the effectiveness of each component. Through THRONE, we take a step forward in devising a data, task, hyperbolicity and network agnostic method for learning representations.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Shivam Agarwal

Owning Collections