User Preference-aware Fake News Detection

Yingtong Dou, Kai Shu, Congying Xia, Philip S. Yu, Lichao Sun

TL;DR

This paper introduces UPFD, a novel framework for fake news detection that goes beyond analyzing news content and propagation patterns. It uniquely incorporates user preference, derived from their historical social media posts, to better identify deceptive content. This approach significantly improves detection accuracy by understanding why users might share certain information, aligning with psychological theories of confirmation bias.

What Problem Does This Paper Address?

The proliferation of fake news on social media poses a significant threat to individuals and society. While existing fake news detection methods primarily focus on analyzing the content of news articles and their spread patterns (exogenous context), they often overlook a crucial factor: the user's own preferences and beliefs (endogenous preference). Psychological theories like confirmation bias suggest that people are more likely to share information that aligns with their existing views. This paper addresses the gap in current research by proposing a method to explicitly model and leverage this user preference to enhance fake news detection.

Methods

The proposed framework, UPFD (User Preference-aware Fake News Detection), integrates both endogenous user preference and exogenous social context.

  • Endogenous Preference Encoding: User preference is captured by analyzing their historical social media posts. The authors collect recent tweets from users who have engaged with a particular news piece. These tweets, along with the news content itself, are encoded using advanced text representation learning models like pre-trained word2vec and BERT. This process aims to infer a user's underlying beliefs and biases.
  • Exogenous Context Extraction: The social context is modeled as a news propagation graph. This graph is constructed based on how users retweet a news piece on platforms like Twitter. The news piece forms the root, and subsequent users who share it become nodes, creating a tree-like structure that illustrates the spread.
  • Information Fusion: A hierarchical approach is used to fuse the endogenous and exogenous information. Graph Neural Networks (GNNs), specifically GraphSAGE, are employed to learn a "user engagement embedding" by aggregating information from the propagation graph. This embedding, along with the news textual embedding (derived from BERT or word2vec), is then concatenated to form the final news embedding. This fused embedding is fed into a neural classifier to predict whether the news is fake or real.

Datasets

The primary dataset used in this research is FakeNewsNet. This dataset is notable because it includes not only news content and fact-checking information but also social engagement data from Twitter, which is crucial for constructing propagation graphs and collecting user historical posts. The paper also mentions Politifact and Gossipcop as datasets for evaluation, though the primary focus for collecting user data is FakeNewsNet.

Key Insights and Ideas

  • User preference is a powerful signal: The confirmation bias theory is a key motivator. The paper demonstrates that understanding a user's existing beliefs (inferred from their past posts) significantly improves fake news detection, even when content and propagation patterns are considered.
  • Endogenous vs. Exogenous Importance: The study reveals that while both endogenous preference and exogenous context are valuable, endogenous information contributes more to performance gain than exogenous information, especially on datasets like Politifact. This highlights the critical role of individual user psychology in information dissemination.
  • BERT's Efficacy: The research confirms that BERT, a powerful pre-trained language model, is more effective than word2vec for encoding textual information, leading to better fake news detection performance.
  • GNNs for Social Context: Graph Neural Networks are well-suited for modeling the complex relationships within social media propagation graphs, effectively capturing the spread dynamics of news.

Key Results

The proposed UPFD framework achieved state-of-the-art performance on the evaluated datasets.

  • UPFD outperformed the best baseline (GCNFN) by approximately 1% in accuracy and F1 score on both Politifact and Gossipcop datasets, with statistical significance.
  • Ablation studies showed that removing either the endogenous preference component (-END) or the exogenous context component (-EXO) from UPFD led to a significant drop in performance. Jointly modeling both yielded the best results.
  • When comparing different text encoders, BERT consistently outperformed word2vec for encoding user preferences and news content.
  • The study also found that using user historical posts (BERT and word2vec features) as node features for GNNs was superior to using only user profile information.

Limitations and Future Work

  • Inaccessible Users: The paper acknowledges a practical challenge: some user accounts might be suspended or deleted. To mitigate this, they used randomly sampled tweets from accessible users engaging with the same news. However, this is an approximation.
  • BERT's Input Length: Due to BERT's input sequence length limitations, encoding 200 historical tweets as a single sequence was not feasible. The authors resorted to encoding each tweet separately and averaging the results, which might not capture the full sequential context of a user's posting history.
  • Future Work: The authors suggest that BERT performance could be further improved through fine-tuning. They also imply that exploring other social media platforms beyond Twitter and investigating more sophisticated GNN architectures could be fruitful avenues for future research.

Summary updated

BibTeX
@inproceedings{dou2021user,
  author = {Yingtong Dou and Kai Shu and Congying Xia and Philip S. Yu and Lichao Sun},
  title = {{User Preference-aware Fake News Detection}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2021},
  url = {https://openreview.net/forum?id=thNEOqoNzEn},
}

Loading related content...

Figure