Calibratable Hetero-NodeRank for measuring node influence

TL;DR

This paper introduces Hetero-NodeRank, a calibratable PageRank-based model that considers network heterogeneity and node attributes to improve influence measurement, validated on real city network data.

Contribution

It proposes a novel, flexible framework that incorporates heterogeneity and calibration, enhancing node influence estimation beyond traditional PageRank methods.

Findings

Outperforms existing algorithms on real city network data

Incorporates heterogeneity and node attributes effectively

Transforms influence measurement into a supervised, calibratable task

Abstract

Node influence metrics have been applied to many applications, including ranking web pages on internet, or locations on spatial networks. PageRank is a popular and effective algorithm for estimating node influence. However, conventional PageRank method considers neither the heterogeneity of network structures nor additional network information, causing a major impedance to performance improvement and an underestimation of non-hub nodes' importance. As these problems are only partially studied, existing solutions are still not satisfying. This paper addresses the problems by presenting a general PageRank-based model framework, dubbed Hetero-NodeRank, that accounts for heterogeneous network topology and incorporates node attribute information to capture both link- and node-based effects in measuring node influence. Moreover, the framework enables the calibration of the proposed model against empirical data, which transforms the original deductive approach into an inductive one that could be useful for hypothesis testing and causal-effect analysis. As the original unsupervised task becomes a supervised one, optimization methods can be leveraged for model calibrations. Experiments on real data from the national city network of China demonstrated that the proposed model outperforms several widely used algorithms.