Core Concepts
Learning invariant representation on non-homophilous graphs is crucial to address distribution shifts and biases.
Abstract
The paper introduces the concept of distribution shifts on non-homophilous graphs, highlighting the limitations of existing methods based on homophily assumptions. It proposes the Invariant Neighborhood Pattern Learning (INPL) framework to address these issues by introducing Adaptive Neighborhood Propagation (ANP) and Invariant Non-Homophilous Graph Learning (INHGL) modules. The ANP module captures adaptive neighborhood information, while the INHGL module learns invariant graph representations. Experimental results demonstrate that INPL outperforms state-of-the-art methods in learning on large non-homophilous graphs.
Stats
Extensive experimental results show that INPL achieves state-of-the-art performance.
The proposed framework addresses bias problems caused by distribution shifts on non-homophilous graphs.
Quotes
"Most nodes are in mixing patterns rather than homophilous or heterophilic patterns."
"We propose a novel bias problem caused by neighborhood pattern distribution shifts on non-homophilous graphs."
"Our contributions include studying a novel bias problem and designing a scalable framework to alleviate unknown distribution shifts."