Fair and robust graph mining

Wang, Yian

Permalink

https://hdl.handle.net/2142/120269 Copy

Description

Title

Fair and robust graph mining

Author(s)

Wang, Yian

Issue Date

2023-04-17

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

Tong, Hanghang

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)

• Fair graph mining
• robustness
• individual fairness
• multi-objective optimization

Abstract

In a connected world, fair graph learning is becoming increasingly important because of the growing concerns about bias. Yet, the vast majority of existing works assume that the input graph comes from a single view while ignoring the multi-view essence of graphs. Generally speaking, the bias in graph mining is often rooted in the input graph and is further introduced or even amplified by the graph mining model. It thus poses critical research questions regarding the intrinsic relationships of fairness on different views and the possibility of mitigating bias on multiple views simultaneously. To answer these questions, in the first part of this thesis , we explore individual fairness in multi-view graph mining. We first demonstrate the necessity of fair multi-view graph learning. Building upon the optimization perspective of fair single-view graph mining, we then formulate our problem as a linear weighted optimization problem. In order to figure out the weight of each view, we resort to the minimax Pareto fairness, which is closely related to the Rawlsian difference principle, and propose an effective solver named iFiG that minimizes the utility loss while promoting individual fairness for each view with two different instantiations. On the other hand, graph neural networks (GNNs) have emerged as a powerful tool for learning representations of graph-structured data, and have achieved state-of-the-art perfor- mance on many tasks such as node classification, link prediction, and graph classification. However, GNNs are also vulnerable to various types of attacks and adversarial settings, such as node deletion, node addition, and edge perturbation. To address this problem, in the second part of the thesis we propose a robust GNN framework based on adversarial train- ing. Our proposed method can learn a robust GNN that can withstand various types of attacks while maintaining high performance on the original task. We show that our pro- posed method can achieve state-of-the-art performance on various benchmark datasets under different attack scenarios. Overall, this thesis contributes to the development of trustworthy graph learning systems that are fair or robust. Our proposed methods have shown promising results on various real-world datasets and benchmark datasets, and can potentially be applied to various do- mains such as social science, healthcare, and cybersecurity. The future work includes further improving the robustness, fairness, and interpretability of graph learning methods, and de- veloping more adaptive and flexible models that can capture the complex dynamics and interactions in real-world networks.

Graduation Semester

2023-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Yian Wang

Owning Collections