A Decision-Based Heterogenous Graph Attention Network for Multi-Class Fake News Detection

TL;DR

This work tackles multiclass fake news detection under semi-supervised constraints by modeling news items as a heterogeneous graph and enabling per-node, per-layer adaptive neighborhood selection. The core method, DHGAT, comprises a decision network that uses $GATv2$ to produce neighborhood-type probabilities $\boldsymbol{\rho}^l_v$ over a set of types $\Gamma$, and a representation network that updates node embeddings using the selected neighborhood type via a Straight-Through $Gumbel$-Softmax sampling $\gamma^l_v$. The final node embeddings are classified with an MLP using a loss that combines cross-entropy over the small labeled set and a semantic-distance term to align predicted scores with ground-truth scores on the six LIAR classes. Evaluated on the LIAR dataset with varying labeled data proportions, DHGAT achieves about a 4% accuracy improvement over strong baselines, demonstrating robustness to limited supervision and the effectiveness of dynamic, task-specific computation graphs for fake news detection.

Abstract

A promising tool for addressing fake news detection is Graph Neural Networks (GNNs). However, most existing GNN-based methods rely on binary classification, categorizing news as either real or fake. Additionally, traditional GNN models use a static neighborhood for each node, making them susceptible to issues like over-squashing. In this paper, we introduce a novel model named Decision-based Heterogeneous Graph Attention Network (DHGAT) for fake news detection in a semi-supervised setting. DHGAT effectively addresses the limitations of traditional GNNs by dynamically optimizing and selecting the neighborhood type for each node in every layer. It represents news data as a heterogeneous graph where nodes (news items) are connected by various types of edges. The architecture of DHGAT consists of a decision network that determines the optimal neighborhood type and a representation network that updates node embeddings based on this selection. As a result, each node learns an optimal and task-specific computational graph, enhancing both the accuracy and efficiency of the fake news detection process. We evaluate DHGAT on the LIAR dataset, a large and challenging dataset for multi-class fake news detection, which includes news items categorized into six classes. Our results demonstrate that DHGAT outperforms existing methods, improving accuracy by approximately 4% and showing robustness with limited labeled data.

Figures (7)

• Figure 1: An overview of the proposed model DHGAT. It takes a heterogeneous graph $HG(V, E, R)$ with various types of edges as input. Initially, using the decision network $\Phi$, which is based on GATv2 DBLP:conf/iclr/Brody0Y22, a probability vector $\rho$ is generated for each node at every layer, where each component represents the probability of a particular neighborhood type $\gamma_i$. Then, the Gumbel-Softmax estimator is used to determine a specific neighborhood type. In the subsequent step, the representation network $\Psi$, also based on GATv2, updates each node's embedding vector based on the determined neighborhood type in the decision network. Finally, the generated embedding vectors are sent to the classification layer, where the final label for each node is predicted.
• Figure 2: Illustration of dynamic neighborhood selection in a heterogeneous graph : (a) In Scenario 1, nodes $A$ and $B$ independently select different edge types ($e_1$ and $e_2$, respectively) to update their embeddings. (b) In Scenario 2, node $A$ selects different edge types across three layers to optimally update its embedding.
• Figure 3: Accuracy of fake news detection using the models: a) GCN and b) GAT, on the graphs constructed based on various speaker profile attributes and the similarity of textual embedding vectors.
• Figure 4: Accuracy of fake news detection using the DHGAT model on heterogeneous graphs constructed with various edge types. For brevity, the edge types in the figure are summarized as follows: sp: speaker, ctx: context, job: job-title, subj: subject.
• Figure 5: Accuracy of fake news detection for each class by the DHGAT model on the heterogeneous graph constructed with two types of edges: speaker and context.