Source-linked AI summary
Hierarchical Propagation Networks for Fake News Detection: Investigation and Exploitation
Kai Shu, Deepak Mahudeswaran, Suhang Wang, Huan Liu
TL;DR
Fake news is hard to detect from content alone, motivating analysis of how stories propagate through hierarchical social-media networks. The paper constructs macro- and micro-level networks, compares linguistic, structural, and temporal features, and tests them for detection. These features consistently outperform existing state-of-the-art methods, are robust across learning algorithms, and achieve average F1 > 0.80.
Problem
Fake news mimics true news and is difficult to detect from content alone, while existing propagation work focuses mainly on macro-level analysis.
Method
The paper constructs macro- and micro-level hierarchical propagation networks, extracts linguistic, structural, and temporal features, compares fake and real news, and evaluates learning algorithms.
Results
Hierarchical propagation features consistently outperform existing state-of-the-art methods, remain robust across learning algorithms, and achieve average F1 > 0.80.
Takeaways & Limitations
Temporal features are more discriminative than linguistic and structural features, while macro- and micro-level features are complementary for fake-news detection.
Takeaways & Limitations
The paper identifies future work in predicting users’ fake-news sharing, unsupervised detection, and combining explicit propagation features with deep learning.
Abstract
from arXiv · showhide
Consuming news from social media is becoming increasingly popular. However, social media also enables the widespread of fake news. Because of its detrimental effects brought by social media, fake news detection has attracted increasing attention. However, the performance of detecting fake news only from news content is generally limited as fake news pieces are written to mimic true news. In the real world, news pieces spread through propagation networks on social media. The news propagation networks usually involve multi-levels. In this paper, we study the challenging problem of investigating and exploiting news hierarchical propagation network on social media for fake news detection. In an attempt to understand the correlations between news propagation networks and fake news, first, we build a hierarchical propagation network from macro-level and micro-level of fake news and true news; second, we perform a comparative analysis of the propagation network features of linguistic, structural and temporal perspectives between fake and real news, which demonstrates the potential of utilizing these features to detect fake news; third, we show the effectiveness of these propagation network features for fake news detection. We further validate the effectiveness of these features from feature important analysis. Altogether, this work presents a data-driven view of hierarchical propagation network and fake news and paves the way towards a healthier online news ecosystem.
1 Introduction
Fake news is difficult to detect from content alone because it is written to mislead readers, while its social-media propagation offers additional structural, temporal, and linguistic signals. This paper investigates hierarchical propagation networks at macro and micro levels and evaluates their usefulness for detection.
- Fake news is intentionally written to mislead readers, making content-only detection nontrivial.
- Social-media propagation has macro-level spreading paths and micro-level conversation trees that provide complementary signals for distinguishing fake from real news.Macro networks capture posts and reposts, whereas micro networks capture replies, sentiment, and stance signals.
- The study asks whether hierarchical propagation features differ between fake and real news and whether they can improve fake-news detection.
- The paper characterizes and statistically compares structural, temporal, and linguistic features across macro- and micro-level networks.
- Hierarchical propagation features consistently outperform existing state-of-the-art methods, achieve average F1 > 0.80 across learning algorithms, and show temporal and structural features outperform linguistic features.
2 Constructing Propagation Networks
The paper constructs hierarchical propagation networks from FakeNewsNet data to represent both global retweet diffusion and local reply conversations. Macro networks infer retweet relationships from user connections and timestamps, while micro networks encode reply relationships.
- The study uses FakeNewsNet, including news content, social context, and dynamic engagement information from GossipCop and PolitiFact.
- The hierarchical network represents information diffusion at micro and macro levels, enabling study of local and global patterns for fake and real news.
- Macro-level nodes represent tweets, and edges represent retweet relationships between them.
- Because official Twitter data does not identify whether a retweet source was an original tweet or another retweet, the method infers likely sources using users’ social connections and retweet timestamps.
- Micro-level nodes represent replies to news tweets, while edges encode direct replies and chains of replies forming conversation threads.
3 Characterizing Propagation Networks
The paper characterizes hierarchical propagation networks at macro and micro levels, comparing fake and real news through structural, temporal, and linguistic features. Across both datasets, several features differ significantly, and propagation features support fake-news classification.
- Macro-level propagation networks: Macro-level networks are analyzed structurally and temporally to characterize global spreading patterns and dissemination dynamics.Because the same textual information is shared across macro-level networks, linguistic analysis is not applied there.
- Macro-level propagation networks: Macro structural features S1, S2, S5, and S7 consistently differ between fake and real news in both datasets.Fake news has significantly greater macro-network depth, indicating longer retweet chains.
- Macro-level propagation networks: Macro temporal features T2, T3, T4, T7, and T8 are statistically significant between fake and real news under t-test.Fake news has shorter lifespans, earlier influential sharing, and shorter posting intervals in the reported datasets.
- Micro-level propagation networks: Micro-level networks represent conversations through structural, temporal, and linguistic features extracted from replies and comment threads.Structural analysis includes conversation-thread patterns, while linguistic analysis examines reply sentiment.
- Micro-level propagation networks: Micro structural features S10, S11, and S14 differ significantly between fake and real news in both datasets.Fake-news micro networks are deeper, and a greater fraction of their cascades contain micro-level reply networks.
- Micro-level propagation networks: Micro temporal features T9, T10, and T11 are statistically significant in both datasets, while fake-news replies show distinct response timing and shorter engagement intervals.Adjacent replies take longer on average, the first reply arrives sooner, and the interval from the first tweet to the last reply is shorter for fake news.
- Micro-level propagation networks: Linguistic features L2 and L3 differ significantly, with replies and direct comments to fake news showing lower average sentiment than those for real news.The reported differences are significant under statistical t-tests in both datasets.
4 Evaluating Propagation Features
The study evaluates hierarchical propagation features for fake-news detection across feature types, network levels, classifiers, and datasets. Combined micro- and macro-level features, especially temporal signals, provide complementary and effective evidence.
- Feature Representation: HPNF concatenates structural, temporal, and linguistic features extracted from each news item's hierarchical propagation network.The network includes macro- and micro-level propagation information.
- Detection Performance: HPNF achieves the best performance on most metrics across both datasets compared with the evaluated baseline representations.The baselines include content-based RST and LIWC, macro-level STFN, and combinations with HPNF.
- Detection Performance: Combining HPNF with existing content or propagation features further improves detection, indicating complementary information across feature spaces.The paper reports improvements for RST HPNF, LIWC HPNF, and STNF HPNF over their component representations.
- Model Robustness: HPNF remains robust across learning algorithms, with average F1 > 0.80 and no significant performance differences across classifiers.Random Forest has the best overall performance, but its advantage over other algorithms is slight.
- Feature Importance Analysis: Combining micro- and macro-level features consistently outperforms either level alone, while the stronger individual level varies by dataset.Micro-level features perform better on PolitiFact, whereas macro-level features perform better on GossipCop.
- Feature Importance Analysis: Temporal features outperform structural and linguistic features on both datasets, and feature-importance patterns identify dataset-specific influential signals.Micro-engagement lifespan is most important on PolitiFact, while retweet-cascade fraction is most important on GossipCop.
5 Related Work
Prior work detects fake news using news content or social context, including user, post, and network signals. This paper addresses limited interpretability and macro-level emphasis by investigating hierarchical propagation features at both micro and macro levels.
- Detection Approaches: Fake-news detection approaches generally use either news content or social context.Content approaches include linguistic and visual features, while social-context approaches use user, post, and network information.
- Detection Approaches: Content-based methods capture writing styles, sensational headlines, and other linguistic or visual cues associated with fake news.Examples include lexical, syntactic, psycholinguistic, and image-based features.
- Detection Approaches: Social-context methods incorporate user characteristics, post responses such as stance, and constructed networks such as diffusion or co-occurrence networks.These signals characterize users and their reactions or relationships around news.
- Research Gap: Existing engagement-based approaches often train classifiers without deeply interpreting the extracted features, creating a black-box problem.The paper positions its investigation as an in-depth analysis of hierarchical propagation networks.
- Research Gap: Prior propagation research mainly analyzes macro-level diffusion, whereas this paper combines macro- and micro-level networks with structural, temporal, and linguistic features for detection.The approach is designed to fill the gap in detailed feature utilization for fake-news classification.
6 Conclusion and Future Work
The paper constructs and analyzes hierarchical propagation networks to determine whether their features distinguish fake from real news and support detection. Experiments find useful, robust, and complementary signals, while future work extends the approach to user prediction, unsupervised detection, and deep learning.
- RQ1: The study constructs macro- and micro-level hierarchical propagation networks and compares structural, temporal, and linguistic features between fake and real news.Statistical analysis evaluates whether these feature distributions differ.
- RQ2: The extracted features significantly contribute to fake-news detection, remain broadly robust across learning algorithms, and show stronger temporal discrimination than linguistic or structural features.Macro- and micro-level features are reported as complementary.
- Future Work: Future work proposes predicting whether users will spread fake news, using hierarchical structure for unsupervised detection, and combining explicit propagation features with deep learning.The proposed directions extend the current supervised feature-based framework.