Understanding Misinformation: The Tale of Fake News and Fake Reviews

Abstract

Misinformation has been long issues in the global communities because of the booming usage of social networks, online retail platforms and so on. The wide spreading of the massive amount of misinformation has recently become a global risk. Therefore, effective detection methods on misinformation is required to combat bad influence. In this dissertation study, we make the following three contributions by focusing on two types of misinformation detection, namely, fake news detection and fake review detection.

The first contribution of this study is the fake news engagement and propagation path framework or FNEPP, in which we devise a novel fake news detection technique from a social-context perspective. The widespread fake news on social media has boosted the demand for reliable fake news detection techniques. Such dissemination of fake news can influence public opinions, allowing unscrupulous parties to control the outcomes of public events such as elections. More recently, a growing number of methods for detecting fake news have been proposed. Most of these approaches, however, have significant limitations in terms of timely detection of fake news. To facilitate early detection of fake news, we propose FNEPP - a unique framework that explicitly combines multiple social context perspectives like news contents, user engagements, user characteristics, and the news propagation path. The FNEPP framework orchestrates two collaborative modules - the engagement module and the propagation path module - as composite features. The engagement module captures news contents and user engagements, whereas the propagation path module learns global and local patterns of user characteristics and news dissemination patterns. The experimental results driven by the two real-world datasets demonstrate the effectiveness and efficiency of the proposed FNEPP framework.

The second contribution of the dissertation lies in an emotion-aware fake review detection framework. Customers are increasingly relying on product reviews when making purchasing decisions. Fake reviews, on the other hand, obstruct the value of online reviews. Thus, automatic fake review detection is required. Previous research devoted most efforts on examining linguistic features, user behavior features, and other auxiliary features in fake review detection. Unfortunately, emotion aspects conveying in the reviews haven’t yet been well explored. After delving in the effective emotion representations mined from review text, we design and implement the emotion-aware fake review detection framework anchored on ensemble learning. The empirical study on the two real-world datasets confirms our model's performance on fake review detection. To investigate how people perceive fake and real reviews differently in terms of emotion aspects, we prepare 200 real product reviews and 200 fake reviews, and random assign 20 reviews to each participant to determine the level of authenticity, credibility, and believability based on 1 - 100 scale. The results from an LIWC-22 emotion analysis intuitively demonstrate people's perception on fake reviews from the aspect of emotions.

The last contribution of the dissertation study is a two-tier text network analysis framework. As the global COVID-19 pandemic boosted the demand of online shopping, the number of online reviews increased dramatically on online shopping platforms. More often than not, customers have the tendency of referring to the product reviews before making buying decisions when products are not physically presented. Fake reviews are designed to influence buyers' purchasing decisions. Existing research devoted their efforts on designing automatic fake review detection systems; however, a text network analysis on fake reviews is missing. To close this technological gap, we construct a two-tier text network analysis framework guiding the investigation of the network-level characteristics and text characteristics of fake reviews. We conduct the extensive experiments driven by the Amazon product review dataset using Gephi. We unfold key findings on guiding the design of next-generation fake-review detection systems.