Source-linked AI summary
MVAN: Multi-View Attention Networks for Fake News Detection on Social Media
Shiwen Ni, Jiawen Li, Hung-Yu Kao
TL;DR
Fake news spreads rapidly on social media, and existing approaches may rely on limited textual or handcrafted information. MVAN combines text semantic attention with propagation structure attention to detect fake news from source tweets and retweet-user structures. On two public datasets, it achieves strong performance, early-detection capability, and reasonable interpretability, while future work plans to add user replies.
Problem
Fake news spreads rapidly on social media, motivating detection from source tweets and retweet users without relying on user comments.
Method
MVAN combines text semantic attention and propagation structure attention to capture clues from source-tweet text and retweet-user propagation graphs.
Results
MVAN significantly outperforms state-of-the-art models in accuracy on two real-world datasets and provides reasonable interpretability and early detection.
Takeaways & Limitations
MVAN provides explanations through key clue words in fake-news text and suspicious users in propagation structure while supporting early detection.
Takeaways & Limitations
Future work will add users’ reply information and model the conversation structure formed by source tweets and replies.
Abstract
from arXiv · showhide
Fake news on social media is a widespread and serious problem in today's society. Existing fake news detection methods focus on finding clues from Long text content, such as original news articles and user comments. This paper solves the problem of fake news detection in more realistic scenarios. Only source shot-text tweet and its retweet users are provided without user comments. We develop a novel neural network based model, \textbf{M}ulti-\textbf{V}iew \textbf{A}ttention \textbf{N}etworks (MVAN) to detect fake news and provide explanations on social media. The MVAN model includes text semantic attention and propagation structure attention, which ensures that our model can capture information and clues both of source tweet content and propagation structure. In addition, the two attention mechanisms in the model can find key clue words in fake news texts and suspicious users in the propagation structure. We conduct experiments on two real-world datasets, and the results demonstrate that MVAN can significantly outperform state-of-the-art methods by 2.5\% in accuracy on average, and produce a reasonable explanation.
Introduction
Fake news spreads rapidly on social media, while existing detection methods are limited by handcrafted features and sequence-only processing. MVAN addresses these gaps by combining graph-based propagation modeling with multiple attention mechanisms and achieves strong accuracy, robustness, and interpretability.
- Motivation: Fake news spreads rapidly online and is difficult for ordinary users to distinguish from abundant social-media information.The paper motivates automated early detection as an auxiliary response to this difficulty.
- Limitations of prior work: Feature-based methods are time-consuming, labour-intensive, and highly dependent on handcrafted-feature quality.The paper states that their performance is not ideal in most cases.
- Limitations of prior work: Sequence-based neural methods process sequential data but cannot process structured data such as propagation graphs.Social-media dissemination can be represented as a graph of retweet users.
- MVAN approach: MVAN combines RNN and graph neural network processing with text semantic attention and propagation structure attention.The design targets information in both source-tweet text and structured propagation.
- Contributions: Graph attention networks are adopted to encode and represent the propagation structure of news.The paper identifies this as a contribution of MVAN.
- Results and interpretability: MVAN achieves the highest accuracy on two real-world datasets, while providing early-detection robustness and interpretability from text and propagation perspectives.The paper reports these properties as experimental and model contributions.
Related Work
Related work spans content-based, handcrafted-feature, sequence, and structure-based fake-news detection methods. The paper distinguishes MVAN by targeting source text without response comments while integrating structural information, explainability, and multiple attention mechanisms.
- Scope: Fake-news detection aims to distinguish the authenticity of social-media news using relevant information such as text content, comments, and propagation structure.The related literature is organized into categories based on these information sources and modeling choices.
- Feature-based methods: Early handcrafted-feature methods extracted signals from text, publisher or user profiles, communication modes, and propagation structures.Examples include decision-tree and feature-enrichment approaches for Twitter and Sina Weibo.
- Content-based methods: Content-based methods use text with RNNs, CNNs, attention, LSTMs, or hierarchical sequence models, with some approaches incorporating comments.These methods primarily model textual or post-level information.
- Structure-based methods: Structure-based methods model conversational, propagation, multimodal social, or follow-follower graphs using recursive networks, GCNs, or GNNs.The cited approaches introduce structural information into fake-news detection.
- MVAN distinction: MVAN targets source news text without response comments and integrates structural information, model explainability, and multiple attention mechanisms.The paper presents these properties as its distinguishing combination relative to related studies.
The background of the related deep learning
The background introduces deep-learning representations for sequential and textual data, then explains attention and graph attention networks as mechanisms for selecting informative information over sequences and graph neighborhoods.
- Deep learning: Deep learning automatically generates vector representations for words, phrases, and sentences instead of relying on manually extracted features.The section frames these techniques as the deep-learning components used in the paper.
- Recurrent neural networks: RNNs model variable-length sequential information by updating hidden states from previous states and current inputs.The equations describe recurrent hidden-state updates over time steps.
- Recurrent neural networks: Basic RNNs cannot learn long-distance temporal dependencies reliably because of vanishing or exploding gradients.LSTM and GRU units are introduced as extensions with memory or gating mechanisms.
- Word embeddings: Word embeddings represent glossary words as continuous vectors so their relationships can be reflected more effectively than with one-hot encoding.Word2vec is presented as an application including CBOW and skip-gram models.
- Attention mechanisms: Attention mechanisms let neural models focus on information considered important, such as more informative words in a sentence.The section describes their use in RNN/CNN-based NLP and self-attention representations.
- Graph attention networks: GATs incorporate attention into graph propagation by computing each node’s hidden state while attending to neighboring nodes.The layer uses node-pair attention coefficients, a shared linear transformation, and a feedforward-network weight vector.
Problem Statement
The model classifies source tweets as true or fake using each tweet’s propagation graph and retweet-user features, while highlighting influential words and users for interpretation.
- MVAN classifies each source tweet as ‘true’ or ‘fake’ using its corresponding propagation graph and retweet users’ feature vectors.
- The model is trained from labelled mappings between source tweets and their true or fake labels.
- MVAN highlights clue words in source tweets and users in the propagation structure to interpret its predictions.
The proposed MVAN model
MVAN combines attention-based representations of source-tweet text and propagation structure, then predicts the news label from both views.
- MVAN uses text semantic attention, propagation structure attention, and a prediction module to detect fake news from source tweets and propagation structures.The prediction module concatenates the two representations before generating the final detection result.
- Text semantic attention networks: Text semantic attention uses word embeddings and a BiGRU to represent source-tweet semantics while assigning weights to words.The resulting vector contains text semantics and each word’s attention weight.
- Propagation structure attention networks: Propagation structure attention transforms user features and applies masked self-attention over each user node’s first-order neighbours.The mechanism includes the node itself as a neighbour and produces updated user-node features.
- Propagation structure attention networks: The propagation-attention output layer replaces concatenation with averaging and uses ReLU as its activation function.
- Propagation structure attention networks: Propagation structure attention produces a representation of the news propagation structure together with attention weights for user nodes and their neighbours.
- Prediction module: The prediction module is a multilayer feedforward network that applies softmax to the combined text and propagation representations to predict the Twitter-news label.Training minimizes cross-entropy loss between predicted and real labels.
Experiment and Results
MVAN is evaluated on Twitter15 and Twitter16 using source tweets, retweet-user information, and standard classification metrics. It outperforms comparison models, while ablations and early-detection analyses examine the contribution, robustness, and interpretability of its two attention mechanisms.
- Experimental Setup: MVAN is compared with machine-learning, recurrent, convolutional, graph, and multi-view neural models using Accuracy, Precision, Recall, and F1.The evaluation uses a 70% training and 30% testing split, with results averaged over ten runs.
- Main Results: 3.06% and 2.03% accuracy gains over G-SEGA are reported on Twitter15 and Twitter16, respectively, while gains over SVM-BOW are about 25% on both datasets.The paper reports significant improvements across all evaluation criteria on both public datasets.
- Ablation Study: About 1% performance is lost when either text semantic attention or propagation structure attention is removed.The ablations retain both information sources but remove one attention mechanism, indicating that each mechanism contributes to performance.
- Ablation Study: Using text attention alone lowers performance by 2.9% to 3.6%, whereas using propagation attention alone lowers it by about 9% on both datasets.Propagation-only models still achieve 8.31% and 8.45% on Twitter15 and Twitter16, respectively, showing detectable clues in propagation structure.
- Early Detection Performance: MVAN reaches approximately 91% accuracy at the earliest stage, with a stable curve indicating robustness in early fake news detection.The early-detection comparison includes MVAN variants using ST-GCN and DCRNN modules.
Conclusion
MVAN combines text semantic attention with propagation structure attention to capture clues from source tweet text and retweet propagation. The model also supports interpretation, while reply information remains a target for future improvement.
- MVAN combines text semantic attention and propagation structure attention to capture hidden clues in source tweet text and propagation structure.The two mechanisms address both textual and propagation information simultaneously.
- Attention weights highlight key clue words in source tweets and identify influential users in retweet propagation.Figure 7 visualizes text attention and retweet-order propagation attention, marking the highest-weight user.
- MVAN is evaluated on two public datasets and provides strong performance with reasonable interpretation ability.
- Adding users’ reply information is identified as future work to further improve the model’s performance.The planned extension would model conversation structure formed by source tweets and replies.