Towards a general diffusion-based information quality assessment model

TL;DR

This paper introduces a diffusion-based, interpretable framework for assessing information quality in academic publications, using diffusion features to predict impact with high accuracy and transparency.

Contribution

It presents a novel, domain-agnostic model leveraging diffusion dynamics and a generalized additive model for scalable, interpretable information quality assessment.

Findings

High correlation (0.834) with future citation gains

Up to 95.62% accuracy in predicting high-impact papers

Timeliness and salience are the most robust predictors

Abstract

The rapid and unregulated dissemination of information in the digital era has amplified the global "infodemic," complicating the identification of high quality information. We present a lightweight, interpretable and non-invasive framework for assessing information quality based solely on diffusion dynamics, demonstrated here in the context of academic publications. Using a heterogeneous dataset of 29,264 sciences, technology, engineering, mathematics (STEM) and social science papers from ArnetMiner and OpenAlex, we model the diffusion network of each paper as a set of three theoretically motivated features: diversity, timeliness, and salience. A Generalized Additive Model (GAM) trained on these features achieved Pearson correlations of 0.834 for next-year citation gain and up to 95.62% accuracy in predicting high-impact papers. Feature relevance studies reveal timeliness and salience as the most robust predictors, while diversity offers less stable benefits in the academic setting but may be more informative in social media contexts. The framework's transparency, domain-agnostic design, and minimal feature requirements position it as a scalable tool for global information quality assessment, opening new avenues for moving beyond binary credibility labels toward richer, diffusion-informed evaluation metrics.