Multilevel User Credibility Assessment in Social Networks

2023-09-23

arXiv.org

AI Summary

Existing social media credibility assessment methods rely solely on binary true/false labels, failing to capture the nuanced spectrum of user reliability. Method: We propose a multi-level credibility quantification framework for fine-grained user trustworthiness ranking and veracity classification (true vs. false news). We construct the first multi-tier annotated user credibility dataset, integrating heterogeneous features—including user profiles, tweets, comments, and social graph structures—and design MultiCred, an end-to-end model that jointly encodes textual features (via BERT) and non-textual features (via graph neural networks and MLP), enabling interpretable, progressive decision-making. Contribution/Results: This work introduces and formalizes the multi-level user credibility paradigm—the first of its kind. On our benchmark dataset, MultiCred achieves over 12% improvement in both accuracy and F1-score over state-of-the-art baselines, demonstrating its effectiveness and practicality for rumor mitigation.

Abstract

Online social networks are major platforms for disseminating both real and fake news. Many users, intentionally or unintentionally, spread harmful content, fake news, and rumors in fields such as politics and business. Consequently, numerous studies have been conducted in recent years to assess user credibility. A significant shortcoming of most existing methods is that they categorize users as either real or fake. However, in real-world applications, it is often more desirable to consider several levels of user credibility. Another limitation is that existing approaches only utilize a portion of important features, which reduces their performance. In this paper, due to the lack of an appropriate dataset for multilevel user credibility assessment, we first design a method to collect data suitable for assessing credibility at multiple levels. Then, we develop the MultiCred model, which places users at one of several levels of credibility based on a rich and diverse set of features extracted from users' profiles, tweets, and comments. MultiCred leverages deep language models to analyze textual data and deep neural models to process non-textual features. Our extensive experiments reveal that MultiCred significantly outperforms existing approaches in terms of several accuracy measures.

Problem

• Social Media
• Multi-level Credibility Assessment
• Fake News Detection

Innovation

• MultiCred
• Multi-level Credibility Assessment
• Advanced Language and Neural Network Models