[Paper Review] Rumor Detection on Twitter Using Multiloss Hierarchical BiLSTM with an Attenuation Factor

Yudianto Sujana, Jiawen Li+1 more|arXiv (Cornell University)|Oct 31, 2020

Misinformation and Its Impacts18 references7 citations

TL;DR

This paper proposes a multiloss hierarchical Bidirectional Long Short-Term Memory (BiLSTM) model with an attenuation factor for rumor detection on Twitter. By structuring the network into post-level and event-level BiLSTM modules with separate loss functions and applying an attenuation factor at the post level, the model improves feature learning and reduces training time, achieving superior performance over state-of-the-art methods on two benchmark datasets.

ABSTRACT

Social media platforms such as Twitter have become a breeding ground for unverified information or rumors. These rumors can threaten people's health, endanger the economy, and affect the stability of a country. Many researchers have developed models to classify rumors using traditional machine learning or vanilla deep learning models. However, previous studies on rumor detection have achieved low precision and are time consuming. Inspired by the hierarchical model and multitask learning, a multiloss hierarchical BiLSTM model with an attenuation factor is proposed in this paper. The model is divided into two BiLSTM modules: post level and event level. By means of this hierarchical structure, the model can extract deep in-formation from limited quantities of text. Each module has a loss function that helps to learn bilateral features and reduce the training time. An attenuation fac-tor is added at the post level to increase the accuracy. The results on two rumor datasets demonstrate that our model achieves better performance than that of state-of-the-art machine learning and vanilla deep learning models.

Motivation & Objective

• To address the challenge of low precision and high training time in existing rumor detection models on Twitter.
• To improve feature extraction from limited textual data using a hierarchical deep learning architecture.
• To enhance model training efficiency and accuracy through multitask learning with multiple loss functions.
• To introduce an attenuation factor at the post level to refine feature learning and boost classification performance.
• To validate the model’s effectiveness on real-world rumor detection benchmarks.

Proposed method

• The model employs a two-level hierarchical structure: a post-level BiLSTM and an event-level BiLSTM to process individual tweets and their cascades, respectively.
• Each level uses a dedicated loss function—post-level and event-level loss—to enable multitask learning and improve feature representation.
• An attenuation factor is applied at the post level to down-weight less reliable or noisy features during training, enhancing model robustness.
• The hierarchical design allows the model to capture both local (post-level) and global (event-level) contextual patterns in rumor propagation.
• The model is trained end-to-end using backpropagation with combined multiloss optimization to jointly optimize both levels.
• The architecture is evaluated on two public rumor detection datasets to assess performance and generalization.

Experimental results

Research questions

• RQ1 Can a hierarchical BiLSTM architecture with multiloss training improve rumor detection accuracy compared to vanilla deep learning models?
• RQ2 Does the inclusion of an attenuation factor at the post level enhance model performance and robustness?
• RQ3 Can the proposed model reduce training time while maintaining or improving precision in rumor detection?
• RQ4 How does the model compare to state-of-the-art machine learning and deep learning approaches on standard benchmark datasets?
• RQ5 To what extent does the hierarchical structure improve feature learning from limited textual data in rumor detection?

Key findings

• The proposed multiloss hierarchical BiLSTM with an attenuation factor achieves higher F1-score and precision than state-of-the-art models on two benchmark rumor detection datasets.
• The model reduces training time significantly due to efficient multitask learning with separate loss functions at each level.
• The attenuation factor at the post level contributes to improved accuracy by suppressing noisy or unreliable features during training.
• The hierarchical structure enables better capture of both micro-level (individual tweet) and macro-level (event cascade) patterns in rumor propagation.
• The model outperforms traditional machine learning and standard deep learning baselines in both F1 and accuracy metrics on both datasets.
• Empirical results confirm that the combination of multiloss training and hierarchical modeling leads to superior generalization and robustness in rumor detection.