Multi-view graph neural network for fair representation learning

Date

2024

Authors

Zhang, G.
Yuan, G.
Cheng, D.
He, L.
Bing, R.
Li, J.
Zhang, S.

Editors

Zhang, W.

Advisors

Journal Title

Journal ISSN

Volume Title

Type:

Conference paper

Citation

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2024 / Zhang, W. (ed./s), vol.14963 LNCS, pp.208-223

Statement of Responsibility

Conference Name

8th Asia-Pacific Web and Web-Age Information Management Joint International Conference on Web and Big Data, APWeb-WAIM 2024 (30 Aug 2024 - 1 Sep 2024 : Jinhua)

Abstract

In Graph Neural Networks, connectivity is usually represented by a fixed adjacency matrix, however, such a matrix fails to capture the complex entanglement present in relational data and is prone to the over-squashing and under-reaching issues. In particular, this rigidity may contribute to algorithmic discrimination, thereby undermining the model’s fairness. To address this limitation, we propose a Multi-view Fair Graph Neural Network (MVFGNN) framework for fair representation learning. Specifically, the MVFGNN framework concurrently learns information from the original view, diffusion view, and feature view in an integrated end-to-end manner. These views are then fused based on their structural similarity, offering a multifaceted approach to node representation learning. Subsequently, an adversarial learning method with a gradient penalty is applied to eliminate sensitive information in the learned node representation. Extensive experiments conducted on three real-world datasets demonstrate the exceptional performance of MVFGNN in terms of both fairness detection and prediction accuracy.

School/Discipline

Dissertation Note

Provenance

Description

Access Status

Rights

Copyright 2024 The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. Access Condition Notes: Accepted manuscript available after 01 October 2025

License

Grant ID

Call number

Persistent link to this record