Testing Alignment of Node Attributes with Network Structure Through Label Propagation

TL;DR

This paper introduces a statistical test to evaluate whether node attributes in a network are aligned with the network's connectivity structure, using label propagation and applied to synthetic and real biological data.

Contribution

It develops a novel statistical testing method for assessing attribute-structure alignment in attributed networks, addressing a gap in existing community detection approaches.

Findings

The test can identify attribute-structure dependencies in synthetic networks.

Applied to CyTOF data, it successfully detects markers associated with cell subpopulations.

The method provides empirical p-values indicating the significance of attribute alignment.

Abstract

Attributed network data is becoming increasingly common across fields, as we are often equipped with information about nodes in addition to their pairwise connectivity patterns. This extra information can manifest as a classification, or as a multidimensional vector of features. Recently developed methods that seek to extend community detection approaches to attributed networks have explored how to most effectively combine connectivity and attribute information to identify quality communities. These methods often rely on some assumption of the dependency relationships between attributes and connectivity. In this work, we seek to develop a statistical test to assess whether node attributes align with network connectivity. The objective is to quantitatively evaluate whether nodes with similar connectivity patterns also have similar attributes. To address this problem, we use a node sampling and label propagation approach. We apply our method to several synthetic examples that explore how network structure and attribute characteristics affect the empirical p-value computed by our method. Finally, we apply the test to a network generated from a single cell mass cytometry (CyTOF) dataset and show that our test can identify markers associated with distinct sub populations of single cells.