The paper proposes CAFIN, a centrality-aware fairness-inducing framework for unsupervised graph representation learning. The key insights are:
Unsupervised graph representation learning frameworks like GraphSAGE tend to focus more on high-degree (popular) nodes, leading to a performance disparity between popular and unpopular nodes.
CAFIN addresses this issue by augmenting the loss function of GraphSAGE to incorporate centrality information and induce fairness constraints. It prioritizes the information flow for less central nodes, leading to more equitable representations.
CAFIN is evaluated on various datasets spanning different domains (citation networks, social networks, co-purchase networks, biological networks). It consistently reduces the performance disparity across nodes (18-80% reduction) while incurring only a minimal cost in overall accuracy.
The paper also explores an approximate distance measure variant of CAFIN to address the high preprocessing time complexity, making it more scalable for larger graphs.
Extensive ablation studies are conducted to analyze the impact of different components of the proposed loss function design and the effectiveness of the approximate distance measures.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Arvindh Arun... at arxiv.org 04-23-2024
https://arxiv.org/pdf/2304.04391.pdfDeeper Inquiries