Source-linked AI summary
User Preference-aware Fake News Detection
Yingtong Dou, Kai Shu, Congying Xia, Philip S. Yu, Lichao Sun
TL;DR
Existing fake news detectors often overlook users’ endogenous preferences, despite historical posts offering information about those preferences. UPFD jointly models historical-post preferences, news content, and propagation graphs; experiments report improved detection performance, while the approach relies on historical-post information that prior datasets lacked.
Problem
Existing fake news detection methods mainly model news content or exogenous context, while work exploiting users’ endogenous preferences remains limited.
Method
UPFD encodes user historical posts and news content, builds a news propagation graph, and jointly fuses endogenous and exogenous information.
Results
UPFD achieves the best performance among the compared baselines and outperforms GCNFN by around 1% on both datasets with statistical significance.
Takeaways & Limitations
Modeling user endogenous preference provides additional information for fake news detection, particularly when user comment information is limited.
Takeaways & Limitations
Prior fake news datasets lacked historical-post information, so the study uses crawled user posts from FakeNewsNet accounts.
Abstract
from arXiv · showhide
Disinformation and fake news have posed detrimental effects on individuals and society in recent years, attracting broad attention to fake news detection. The majority of existing fake news detection algorithms focus on mining news content and/or the surrounding exogenous context for discovering deceptive signals; while the endogenous preference of a user when he/she decides to spread a piece of fake news or not is ignored. The confirmation bias theory has indicated that a user is more likely to spread a piece of fake news when it confirms his/her existing beliefs/preferences. Users' historical, social engagements such as posts provide rich information about users' preferences toward news and have great potential to advance fake news detection. However, the work on exploring user preference for fake news detection is somewhat limited. Therefore, in this paper, we study the novel problem of exploiting user preference for fake news detection. We propose a new framework, UPFD, which simultaneously captures various signals from user preferences by joint content and graph modeling. Experimental results on real-world datasets demonstrate the effectiveness of the proposed framework. We release our code and data as a benchmark for GNN-based fake news detection: https://github.com/safe-graph/GNN-FakeNews.
1 INTRODUCTION
Fake news detection has largely modeled news content and exogenous social context while overlooking users’ endogenous preferences. UPFD addresses this gap by jointly modeling historical-post preferences and propagation context.
- Social-media disinformation has harmful effects and creates demand for improved fake news detection.
- Existing approaches use news content, engineered features, deep learning, or propagation graphs but ignore user endogenous preferences.
- Confirmation bias suggests users prefer information that confirms their existing views, motivating preference-aware detection.
- UPFD models endogenous preference from user historical posts while jointly incorporating exogenous context.
- The paper studies user preference-aware fake news detection and proposes exploiting endogenous preference with exogenous context.
2 OUR APPROACH
UPFD encodes user preferences from historical posts and exogenous context from Twitter propagation graphs, then fuses both with news content for credibility prediction.
- Endogenous Preference Encoding: UPFD learns users’ endogenous preferences by encoding their historical posts and encodes the news text using text representation methods.The approach uses pretrained word2vec or BERT representations.
- Endogenous Preference Encoding: The method uses FakeNewsNet and crawls historical tweets from accounts that retweeted each news item.It collects the recent two hundred tweets for each account, with inaccessible users handled using sampled tweets from accessible engaging users.
- Exogenous Context Extraction: UPFD constructs a tree-structured propagation graph whose root is the news item and whose other nodes are users sharing it.Retweet timing and follow relationships determine estimated propagation paths.
- Information Fusion: A GNN aggregates node features over the propagation graph, and mean pooling produces a user engagement embedding.
- Information Fusion: UPFD concatenates the user engagement embedding with the news textual embedding and classifies credibility with a two-layer MLP.Training uses binary cross-entropy and SGD.
3 EXPERIMENTS
The experiments evaluate UPFD against prior methods and examine how endogenous and exogenous information contribute to performance.
- The experiments address performance comparisons with previous works and the contributions of endogenous, exogenous, and other framework variants.
3.1 Experimental Setup
Experiments use FakeNewsNet and compare UPFD with baselines that encode different combinations of news content, user information, and propagation graphs under unified training settings.
- Dataset: FakeNewsNet contains fake and real news from two fact-checking websites together with related Twitter social-engagement information.
- Baselines: The baseline comparison covers content-only models, propagation-graph models, and models using user comments or profile attributes.The baselines include CSI, SAFE, GCNFN, GNN-CL, and MLP variants using word2vec or BERT news embeddings.
- Implementation: All models use a unified graph embedding size, batch size, optimizer, regularization weight, and 20%-10%-70% train-validation-test split.Results are averaged over five runs.
3.2 RQ1: Performance Evaluation
UPFD achieves the best fake news detection performance among the evaluated baselines, outperforming GCNFN by about 1% on both datasets with statistical significance. The results indicate that historical user posts provide additional information beyond news content, user comments, and propagation context.
- UPFD achieves the best fake news detection performance among the six evaluated baselines.Table 2 reports the comparison between UPFD and six baselines.
- Around 1%, UPFD outperforms the best baseline, GCNFN, on both datasets with statistical significance.
- Historical user posts improve fake news detection because all baselines omit these endogenous preference signals.
- The best-performing UPFD configuration on both datasets uses BERT as the text encoder and GraphSAGE as the graph encoder.
3.3 RQ2: Ablation Study
The ablation study evaluates endogenous user preferences, exogenous propagation context, and their joint modeling through encoder and framework variants. Endogenous features outperform profile features, while removing either information source reduces performance.
- 3.3.1 Encoder Variants: Endogenous word2vec and BERT features consistently outperform profile features that encode only user profiles.
- 3.3.1 Encoder Variants: GraphSAGE and BERT have the average best performance among the evaluated model and feature variants.
- 3.3.2 Framework Variants: The ablation fixes the text and graph encoders while removing endogenous information, exogenous information, or both.
4 CONCLUSION
UPFD models users’ historical posts as endogenous news-consumption preferences and combines them with news propagation graphs as exogenous social context. Experiments demonstrate the advantage of modeling endogenous user preference for fake news detection.
- UPFD uses users’ historical posts to model endogenous preference and news propagation graphs to represent exogenous social context.
- Experimental results demonstrate the advantage of modeling endogenous user preference for fake news detection.