Rumor detection in social media based on a hierarchical attention network

Xueqi Cheng, Xiangwen Liao, Zhi Huang, Guolong Chen, Dingda YangScientia 2018 · DOI

Identifying rumors on social media is a critical challenge in the digital age. While deep learning has significantly advanced detection capabilities, many existing models fail to capture the nuanced temporal dynamics of online conversations. Researchers from Fuzhou University and the Chinese Academy of Sciences have proposed a novel Hierarchical Attention Network (HAN) that addresses these gaps by analyzing the "shape" of a conversation at both micro and macro levels.

The Challenge: Dynamics and Latent Signals

Rumor detection is a race against time. Existing models often suffer from two primary limitations:

  1. Ignoring Intra-interval Dynamics: Many models partition events into time intervals but treat the microblogs within those windows as an unordered "bag of words," losing the chronological back-and-forth that characterizes rumor propagation.
  2. Neglecting Latent Signals: Modern deep learning often favors raw text embeddings over "expert knowledge." However, traditional feature engineering—such as identifying questioning language ("Is this true?") or user credibility—remains a powerful indicator of misinformation.

The HAN-FC Architecture

The proposed model, HAN-FC (Hierarchical Attention Network with Feature Combination), utilizes a three-module architecture to bridge the gap between deep temporal learning and hand-crafted features.

The architecture of the proposed Hierarchical Attention Network (HAN-FC), illustrating the flow from microblog embedding to event classification across three distinct modules.

Module 1: Microblog-level Representation

The event is divided into discrete time intervals. Within each interval, a sequence of microblogs is embedded and processed through a Bidirectional Gated Recurrent Unit (BiGRU). An attention mechanism assigns weights to individual posts, identifying high-signal content—such as a debunking tweet—amidst the noise.

Module 2: Feature Fusion

The learned representation from the first module is concatenated with hand-crafted features. These include content-based signals (regex-based detection of inquiries and corrections) and user-based features (account age, follower ratios, and verification status).

Module 3: Time Interval-level Representation

The fused vectors for all intervals are fed into a second BiGRU with its own attention layer. This captures the "macro" evolution of the event, tracking how the conversation shifts from initial shock to skepticism or confirmation over time.

Experimental Results and Insights

The researchers evaluated HAN-FC on two major benchmarks: Sina Weibo (4,664 events) and Twitter (992 events). The model achieved superior results across the board:

  • Performance: HAN-FC reached 96.8% accuracy on Sina Weibo and 78.7% on Twitter, outperforming previous state-of-the-art models (like CSI) by approximately 1.5%.
  • Early Detection: On the Weibo dataset, the model reached stable, high accuracy within just 12 hours of the first tweet. By learning the specific patterns of how rumors emerge, the model can make accurate calls significantly faster than traditional SVM-based methods.
  • The Power of "Questioning": One of the most effective features was the detection of "Enquiries and Corrections." Rumors often trigger specific linguistic responses that raw neural networks might overlook but simple pattern matching captures effectively.

Conclusion and Future Directions

By combining hierarchical deep learning with traditional feature engineering, HAN-FC provides a robust framework for early rumor detection. Future work aims to address data completeness issues—such as when tweets are deleted—and explore the use of Reinforcement Learning to dynamically partition time windows based on conversation density rather than fixed intervals.

Summary updated

Abstract

For rumor detection in social media, the majority of feature-representation-based studies capture textual features or global users features according to a partitioned sequence of microblogs. However, these studies ignore the time-series information across the microblogs in one time interval. Moreover, the latent textual information and the local users information, which have been proven effective according to the traditional machine learning methods, are overlooked in capturing the time intervals features, resulting in low performance. Therefore, we propose a rumor detection method in social media based on a hierarchical attention network. First, microblogs are partitioned into several time intervals. Then, the variation of information across the microblogs changing over time is learned using a bidirectional gated recurrent unit neural network with an attention mechanism. After that, the variation of features across the microblogs is combined with hand-crafted features to incorporate latent textual information and local users information into the time intervals features. Finally, we capture features variation across the time intervals by using the bidirectional gated recurrent unit neural network, with the attention mechanism, and classify microblog events. Experimental results over two public datasets, Sina Weibo and Twitter, show that the proposed method outperforms (in terms of the accuracy) state-of-the-art methods by 1.5% and 1.4% over the two datasets, respectively, and it is effective for rumor detection.
BibTeX
@inproceedings{cheng2018rumor,
  author = {Xueqi Cheng and Xiangwen Liao and Zhi Huang and Guolong Chen and Dingda Yang},
  title = {{Rumor detection in social media based on a hierarchical attention network}},
  booktitle = {Scientia},
  year = {2018},
  doi = {10.1360/n112018-00134},
}
Figure