Fake News Detection on News-Oriented Heterogeneous Information Networks through Hierarchical Graph Attention

Yuxiang Ren, Jiawei ZhangIJCNN 2021 · OpenReview

TL;DR

This paper introduces HGAT, a novel Hierarchical Graph Attention Network designed to detect fake news by modeling news articles, creators, and topics as a heterogeneous information network (HIN). HGAT leverages a hierarchical attention mechanism to learn rich node representations, outperforming existing methods on real-world datasets.

What Problem Does This Paper Address?

The proliferation of fake news poses a significant threat to public discourse and trust. Traditional fake news detection methods often rely solely on textual content, which can be easily manipulated by sophisticated actors who mask writing styles. This paper addresses the limitation of text-centric approaches by proposing a more holistic method that considers the interconnectedness of news articles with their creators and associated topics. These entities and their relationships can be effectively represented as a Heterogeneous Information Network (HIN), where news articles are nodes with various types of connections to other nodes. The challenge lies in effectively learning representations from such complex, multi-typed networks to accurately classify fake news.

Methods

The core of the proposed solution is the Hierarchical Graph Attention Network (HGAT). HGAT tackles the heterogeneity of HINs through a two-level attention mechanism:

  • Node-Level Attention: This mechanism learns the importance of neighbors of the same type for a given news article node. For instance, it determines how much attention a news article should pay to other news articles written by the same creator, or to other articles related to the same subject. This is achieved by projecting node features into a common space and then applying an attention mechanism. The aggregated representations of same-typed neighbors form "schema nodes."
  • Schema-Level Attention: This layer then learns the importance of these different "schema nodes" (representing aggregated information from different neighbor types) for the target news article. This allows HGAT to dynamically weigh the influence of creators, subjects, and other news articles when forming the final representation of a news article.

This hierarchical approach allows HGAT to capture nuanced relationships within the HIN and learn more discriminative representations for fake news detection. The model is trained end-to-end using a cross-entropy loss function for classification.

Datasets

The authors evaluate HGAT on two real-world datasets:

  • PolitiFact: This dataset is derived from the PolitiFact fact-checking website and includes news articles, their creators, and associated subjects. The fact-checking results provide labels for the news articles, ranging from "True" to "Pants on Fire!". This dataset is used for both binary (fake/real) and multi-class (6 credibility levels) classification.
  • BuzzFeed: This dataset, sourced from Shu et al. [15], contains news articles from BuzzFeed, categorized by their publishers and associated Twitter users. It is used for binary classification (true/fake).

Key Insights and Ideas

  • Beyond Text: The Power of Context: The paper strongly advocates for moving beyond purely text-based analysis. By incorporating information about news creators and topics within a HIN, HGAT demonstrates that contextual information is crucial for accurate fake news detection.
  • Hierarchical Attention for Heterogeneity: The novel hierarchical attention mechanism is key. It effectively handles the diverse node and edge types in HINs by first aggregating information from same-typed neighbors and then learning the importance of these aggregated representations. This is a significant advancement over methods that either treat HINs as homogeneous or rely on handcrafted meta-paths.
  • Meta-path Free Generalizability: HGAT's success without relying on predefined meta-paths (a common requirement for many HIN embedding methods) highlights its flexibility and potential for broader applications on heterogeneous graphs.
  • Fake News is Time-Sensitive: The authors emphasize that past news data may not directly help detect emerging fake news, underscoring the need for models that can learn from the underlying network structure and relationships rather than just static content.

Key Results

  • Superior Performance: HGAT consistently outperforms various baseline models, including text classification methods (SVM, LIWC), network embedding methods (LP, DeepWalk, LINE), and other GNNs (GCN, GAT, HAN), on both PolitiFact and BuzzFeed datasets across multiple metrics (Accuracy, F1, Recall, Precision).
  • Outperforming HAN: HGAT shows an advantage over HAN, another method designed for heterogeneous graphs, suggesting that its hierarchical attention mechanism is more effective than HAN's meta-path-based approach.
  • Effectiveness of Schema-Level Attention: Experiments comparing HGAT with a variant that uses fixed weights for schema nodes (HGAT_No_schema) demonstrate that the learned attention weights significantly improve performance, validating the importance of distinguishing between different neighbor types.
  • Robustness across Training Ratios: HGAT maintains strong performance across different training data ratios, indicating its ability to learn effectively even with limited labeled data.

Limitations and Future Work

  • Textual Feature Simplification: The paper uses TF-IDF for textual feature extraction, which is a relatively simple method. The authors suggest integrating more powerful Natural Language Processing (NLP) techniques to construct a more comprehensive end-to-end model in the future.
  • Broader Applications: While the paper focuses on fake news detection, HGAT is presented as a general graph representation learning model. The authors believe it has significant potential for other node classification tasks on heterogeneous graphs and suggests this as a direction for future research.
  • Exploring Publisher Behavior: The authors mention that mining the behavior patterns of high-risk fake news publishers is a meaningful and interesting avenue for future work.

Summary updated

Abstract

The viral spread of fake news has caused great social harm, making fake news detection an urgent task. Current fake news detection methods rely heavily on text information by learning the extracted news content or writing style of internal knowledge. However, deliberate rumors can mask writing style, bypassing language models and invalidating simple text-based models. In fact, news articles and other related components (such as news creators and news topics) can be modeled as a heterogeneous information network (HIN for short). In this paper, we propose a novel fake news detection framework, namely Hierarchical Graph Attention Network (HGAT), which uses a novel hierarchical attention mechanism to perform node representation learning in HIN, and then detects fake news by classifying news article nodes. Experiments on two realworld fake news datasets show that HGAT can outperform text-based models and other network-based models. In addition, the experiments prove the expandability and generalizability of our for graph representation learning and other node classification related applications in heterogeneous graphs.
BibTeX
@inproceedings{ren2021fake,
  author = {Yuxiang Ren and Jiawei Zhang},
  title = {{Fake News Detection on News-Oriented Heterogeneous Information Networks through Hierarchical Graph Attention}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=s8LpaJpnZDk},
}
Figure