HMBI: A New Hybrid Deep Model Based on Behavior Information for Fake News Detection

Jian Xing,Shupeng Wang,Xiaoyu Zhang,Yu Ding,Zhaolong Ning

doi:10.1155/2021/9076211

Jian Xing, Shupeng Wang + Show 3 more

Open Access

https://doi.org/10.1155/2021/9076211

Copy DOI

Abstract

Fake news can cause widespread and tremendous political and social influence in the real world. The intentional misleading of fake news makes the automatic detection of fake news an important and challenging problem, which has not been well understood at present. Meanwhile, fake news can contain true evidence imitating the true news and present different degrees of falsity, which further aggravates the difficulty of detection. On the other hand, the fake news speaker himself provides rich social behavior information, which provides unprecedented opportunities for advanced fake news detection. In this study, we propose a new hybrid deep model based on behavior information (HMBI), which uses the social behavior information of the speaker to detect fake news more accurately. Specifically, we model news content and social behavior information simultaneously to detect the degrees of falsity of news. The experimental analysis on real-world data shows that the detection accuracy of HMBI is increased by 10.41% on average, which is the highest of the existing model. The detection accuracy of fake news exceeds 50% for the first time.

Highlights

Due to the timeliness and convenience of social media, people are more likely to consume news from social media than traditional news organizations
Through the solution of this problem, we propose a new hybrid deep model based on behavior information for fake news detection, called HMBI
The key contributions of our work are summarized as follows: (1) Model-oriented: we propose a new HBMI model, which is the first time to apply Bidirectional Encoder Representations from Transformers (BERT) [10], Transformer [11], and Convolutional Neural Networks (CNN) [12] synthetically for fake news detection

Summary

Introduction

Due to the timeliness and convenience of social media, people are more likely to consume news from social media than traditional news organizations. Note that few papers existing in the literatures use social behavior information to detect fake news. Our research mainly answers this question: how to effectively use the social behavior information of the speaker to improve the detection accuracy. Through the solution of this problem, we propose a new hybrid deep model based on behavior information for fake news detection, called HMBI. Through the combination of BERT and CNN, we extract more useful sentence embeddings of news content and word embeddings of textual social behavior information. We utilize the speaker’s viewpoint or stance for capturing the speaker’s social engagements and other auxiliary information to learn more useful social behavior features on the Transformer model (3) The experimental analysis on real-world data shows that HMBI is more accurate than previous work on fake news detection.

Related Work

Problem Definition

The Proposed Model

Experiments

Conclusion