Source-linked AI summary
GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media
Yi-Ju Lu, Cheng-Te Li
TL;DR
The paper asks how to detect fake news from short source tweets when user comments and network structures are unavailable, while still producing explanations. It proposes GCAN, which models retweet users, propagation, interactions, and source-text correlations with graph-aware representations and dual co-attention. GCAN outperforms competing methods across two datasets and supports early detection and reasonable explanations.
Problem
The task is fake-news detection from short tweets and retweet-user sequences without user-comment text or social and diffusion network structures, while highlighting explanatory users and words.
Method
GCAN learns source-text, retweet-propagation, and graph-aware user-interaction representations, then uses dual co-attention to model their correlations.
Results
GCAN significantly outperforms the best competing methods across all metrics on two datasets, improving average performance by around 17% on Twitter15 and 15% on Twitter16.
Takeaways & Limitations
GCAN provides fake-news predictions with reasonable explanations and achieves accurate early detection even with only ten retweeters.
Takeaways & Limitations
Some related heterogeneous-information-network methods cannot detect truthfulness for new-coming tweets because they are not inductive.
Abstract
from arXiv · showhide
This paper solves the fake news detection problem under a more realistic scenario on social media. Given the source short-text tweet and the corresponding sequence of retweet users without text comments, we aim at predicting whether the source tweet is fake or not, and generating explanation by highlighting the evidences on suspicious retweeters and the words they concern. We develop a novel neural network-based model, Graph-aware Co-Attention Networks (GCAN), to achieve the goal. Extensive experiments conducted on real tweet datasets exhibit that GCAN can significantly outperform state-of-the-art methods by 16% in accuracy on average. In addition, the case studies also show that GCAN can produce reasonable explanations.
1 Introduction
The paper targets fake-news detection from short source tweets and retweet-user sequences, without user comments or network structures, while requiring explanations. GCAN combines user, propagation, graph-aware interaction, and co-attention representations for detection and explanation.
- Motivation: Social networking supports information exchange and collective intelligence but also enables misinformation such as fake news to propagate.Fake news can mislead the public and produce unjust political, economic, or psychological profit for some parties.
- Challenges: Short tweets create severe data sparsity, while most users reshare stories without comments and retweet diffusion structures are costly to obtain.These constraints limit approaches relying on long documents, rich user comments, or tree-based propagation structures.
- Problem setting: The proposed scenario uses a short source tweet, a retweet-user sequence, and user profiles, excluding user-comment text and social or diffusion network structures.The model must also highlight suspicious retweeters and the source words they attend to as evidence.
- Approach: GCAN learns source-text embeddings, retweet-propagation representations, graph-aware user-interaction representations, and correlations through dual co-attention.Binary prediction is generated from the learned embeddings.
- Contributions: GCAN studies a more realistic social-media fake-news scenario and is designed to learn user interactions, propagation, and their correlation with source text.The paper also presents the dual co-attention mechanism as a source of reasonable explanations.
2 Related Work
Prior work detects fake news through content, user, propagation-structure, or hybrid signals, but the paper emphasizes a combination of short text, no user comments, and explainability.
- Content-based approaches: Content-based approaches commonly use long news text and textual features, including TF-IDF, topics, linguistic properties, writing style, and social emotions.Some methods also use enquiry phrases or recurrent networks to represent user responses.
- User-based approaches: User-based approaches model retweeter traits from account features, profile differences, recurrent-convolutional representations, or heterogeneous graph embeddings.These methods target characteristics such as verification, demographics, follower counts, and shared-account behavior.
- Structure-based approaches: Structure-based approaches exploit propagation patterns or implicit links in social networks to classify misinformation and rumors.Examples include hashtag and URL connections, high-order rumor patterns, and tree-structured propagation representations.
- Limitations of prior work: Some prior heterogeneous-information-network methods cannot detect the truthfulness of new-coming tweets because they do not support the inductive setting.This is identified as a limitation of those related approaches.
- Hybrid-based approaches: Hybrid approaches fuse source content with response comments, user profiles, social interactions, or multimodal information.CSI, event-adversarial models, and dEFEND combine different context sources for detection or explanation.
- Positioning: The paper distinguishes its approach by targeting short text, requiring no user response comments, and allowing model explainability.A related-work comparison organizes studies by news-story text, response comments, user characteristics, propagation structure, social network, and explainability.
3 Problem Statement
The task is binary truthfulness prediction for a source tweet using its retweet propagation path and user features, while also identifying explanatory users and words.
- Data representation: A tweet story is represented as a short-text document containing a sequence of words.The notation defines a set of tweet stories and users.
- Propagation path: Each propagation record contains a retweeting user, that user’s feature vector, and the retweet time.The propagation path contains K retweet records, with the original sharer first and later retweets occurring at later times.
- Prediction target: Each story has a binary truthfulness label, where 0 denotes true and 1 denotes fake.The prediction target is the truthfulness yi of the source story.
- Explainability: The model must highlight a few retweeting users and source-tweet words that explain why the story is classified as true or fake.This makes explanation an explicit requirement alongside binary classification.
4 The Proposed GCAN Model
GCAN combines source-tweet encoding with user-propagation and graph-aware interaction representations, then uses dual co-attention to connect words with retweet users for explainable detection.
- 4 The Proposed GCAN Model: GCAN extracts user characteristics, encodes source-tweet words, models user propagation, represents user interactions with a graph, and applies dual co-attention.Its components cover user features, word embeddings, sequential propagation, graph convolution, and source-user correlation learning.
- 4.1 User Characteristics Extraction: User feature vectors combine profile metadata with retweet-context attributes, including follower counts, verification, timing, and propagation-path length.The paper lists ten features derived from user profiles and retweet behavior.
- 4.2 Source Tweet Encoding: The source tweet is zero-padded to a maximum length and transformed from one-hot word inputs into embeddings before GRU sequence encoding.The word-level encoder produces V = [v1, v2, ..., vm] ∈ R^d×m, followed by GRU-based sequence representation.
- 4.3 User Propagation Representation: Propagation representations use GRU and one-dimensional CNN networks over a fixed-length sequence of retweet-user feature vectors.Longer sequences are truncated to the first n users, while shorter sequences are resampled until they reach length n.
- 4.4 Graph-aware Propagation Representation: GCAN constructs a fully connected user graph whose edge weights are cosine similarities between user feature vectors, then applies stacked GCN layers.The resulting graph-aware representation incorporates neighborhood information from user interactions.
- 4.5 Dual Co-attention Mechanism: Dual co-attention jointly attends to source-tweet words and propagated or interacting users to produce interpretable cross-modal evidence.The mechanism uses proximity matrices and attention probabilities to generate co-attention feature vectors for words and users.
5 Experiments
Experiments on Twitter15 and Twitter16 evaluate GCAN against baselines, test component contributions and early detection, and examine explanations from attention weights. GCAN improves accuracy across varying retweet counts and highlights evidential words and users.
- Datasets and Evaluation Settings: Experiments use Twitter15 and Twitter16 source tweets with corresponding retweet-user sequences, evaluating Accuracy, Precision, Recall, and F1 under repeated 70/30 train-test splits.The datasets retain only true and fake labels, with user information obtained through Twitter API crawling.
- Early Detection: 90% accuracy with only ten retweeters shows that GCAN supports accurate early detection while consistently outperforming competitors as observed retweet counts vary.Accuracy is reported while varying the number of observed retweet users from 10 to 50 in Twitter15 and Twitter16.
- Ablation Analysis: Removing dual co-attention, GRU-based, graph-aware, or CNN-based representations produces ablated GCAN variants for measuring component contributions.Removing the source tweet and attention causes a significant accuracy drop, underscoring the source tweet's role in prediction.
- Explainability: Source-propagation co-attention provides explanations by highlighting source-tweet words and retweet users according to their attention weights.Source-interaction co-attention is excluded from explanations because graph-learned user interaction features are not intuitively interpretable.
- Explainability: GCAN attends to early retweeters and can distribute fake-news evidence across propagation, while case studies identify suspicious user traits and attended words.Highlighted examples include “breaking” and “strict” or “pipeline,” alongside unverified accounts and shorter account descriptions.
6 Conclusion
GCAN detects fake news from short-text tweets and their retweeter sequences while providing explanations. It also supports early detection and may extend to other short-text social-media classification tasks.
- GCAN predicts whether a short-text tweet is fake using the sequence of its retweeters.
- GCAN provides reasonable explanations by identifying evidence associated with suspicious retweeters and the words they concern.
- GCAN can provide early detection of fake news with satisfying performance.
- The authors propose extending GCAN to sentiment detection, hate speech detection, and tweet popularity prediction.