[Paper Review] Adversarial Active Learning based Heterogeneous Graph Neural Network for Fake News Detection

Yuxiang Ren, Bo Wang+2 more|arXiv (Cornell University)|Jan 27, 2021

Misinformation and Its Impacts51 references4 citations

TL;DR

This paper proposes AA-HGNN, a novel adversarial active learning-based heterogeneous graph neural network for fake news detection that leverages a hierarchical attention mechanism to model heterogeneous information networks (HINs) and an adversarial selector to query high-value training samples. The framework achieves state-of-the-art performance with significantly fewer labeled examples, outperforming text-only and standard GNN models on real-world datasets by up to 9% in F1-score under low-data regimes.

ABSTRACT

The explosive growth of fake news along with destructive effects on politics, economy, and public safety has increased the demand for fake news detection. Fake news on social media does not exist independently in the form of an article. Many other entities, such as news creators, news subjects, and so on, exist on social media and have relationships with news articles. Different entities and relationships can be modeled as a heterogeneous information network (HIN). In this paper, we attempt to solve the fake news detection problem with the support of a news-oriented HIN. We propose a novel fake news detection framework, namely Adversarial Active Learning-based Heterogeneous Graph Neural Network (AA-HGNN) which employs a novel hierarchical attention mechanism to perform node representation learning in the HIN. AA-HGNN utilizes an active learning framework to enhance learning performance, especially when facing the paucity of labeled data. An adversarial selector will be trained to query high-value candidates for the active learning framework. When the adversarial active learning is completed, AA-HGNN detects fake news by classifying news article nodes. Experiments with two real-world fake news datasets show that our model can outperform text-based models and other graph-based models when using less labeled data benefiting from the adversarial active learning. As a model with generalizability, AA-HGNN also has the ability to be widely used in other node classification-related applications on heterogeneous graphs.

Motivation & Objective

• To address the challenge of limited labeled data in fake news detection, especially given the high cost and timeliness constraints of expert labeling.
• To model fake news detection as a node classification problem in a news-oriented heterogeneous information network (News-HIN) that integrates creators, subjects, and articles with their relationships.
• To enhance model performance under data scarcity by integrating active learning with adversarial training to select the most informative samples for labeling.
• To develop a generalizable framework that can be applied to diverse node classification tasks on heterogeneous graphs beyond fake news detection.

Proposed method

• The model employs a hierarchical graph attention network (HGAT) to learn node representations by attending over different types of neighbors in the heterogeneous network, preserving both structural and semantic heterogeneity.
• An adversarial active learning framework is introduced, where a selector network is trained in competition with the classifier to identify the most uncertain or informative nodes for labeling.
• The selector uses a differentiable query strategy that maximizes the expected information gain, enabling end-to-end training of the active learning pipeline.
• The framework jointly optimizes the graph neural network for node classification and the selector for query selection, enabling efficient and effective data labeling under scarcity.
• Textual features from news articles are embedded and fused with structural representations from the HIN to enrich node embeddings.
• The model is trained end-to-end using cross-entropy loss on labeled nodes, with the active learning loop iteratively expanding the training set through high-value queries.

Experimental results

Research questions

• RQ1 Can a heterogeneous graph neural network with a hierarchical attention mechanism effectively model multi-type entities and relationships in fake news detection?
• RQ2 How does adversarial active learning improve model performance when labeled data is scarce in fake news detection?
• RQ3 Does the proposed adversarial selector outperform heuristic query strategies (e.g., entropy-based or random selection) in selecting informative samples for labeling?
• RQ4 To what extent does the integration of HINs and active learning reduce the required number of labeled examples while maintaining or improving detection accuracy?
• RQ5 Can the proposed framework generalize to other node classification tasks on heterogeneous graphs beyond fake news detection?

Key findings

• AA-HGNN outperforms text-only models and standard GNNs (e.g., GCN, GAT) in fake news detection, achieving a 9% higher F1-score when trained on only 1,200 labeled nodes compared to 2,800 for the baseline classifier.
• The adversarial active learning strategy significantly improves performance: AA-HGNN achieves a 9% performance gain over the same model without active learning when using 2,800 training nodes.
• The adversarial selector consistently queries higher-quality samples than entropy-based or random selection strategies, as evidenced by superior and stable performance gains across all query batches.
• HGAT-based models outperform homogeneous GNNs (GCN, GAT) on News-HINs, with GCN achieving 0.9688 precision but only 0.0246 recall on the PolitiFact dataset, indicating poor detection of fake news.
• AA-HGNN maintains strong performance even at low data ratios (e.g., 20% of training data), where other models fail to generalize, demonstrating robustness under data scarcity.
• The model generalizes well to other node classification tasks on heterogeneous graphs due to its architecture and training paradigm, suggesting broad applicability beyond fake news detection.