From rumor to genetic mutation detection with explanations: a GAN approach

Mingxi Cheng, Yizhi Li, Shahin Nazarian, Paul BogdanScientific 2021 · DOI

Abstract

Social media have emerged as increasingly popular means and environments for information gathering and propagation. This vigorous growth of social media contributed not only to a pandemic of rumors and misinformation, but also to an urgent need for text-based rumor detection strategies. To speed up the detection of misinformation, traditional rumor detection methods based on hand-crafted feature selection need to be replaced by automatic artificial intelligence (AI) approaches. AI decision making systems require to provide explanations in order to assure users of their trustworthiness. Inspired by the thriving development of generative adversarial networks (GANs) on text applications, we propose a GAN-based layered model for rumor detection with explanations. To demonstrate the universality of the proposed approach, we demonstrate its benefits on a gene classification with mutation detection case study. Similarly to the rumor detection, the gene classification can also be formulated as a text-based classification problem. Unlike fake news detection that needs a previously collected verified news database, our model provides explanations in rumor detection based on tweet-level texts only without referring to a verified news database. The layered structure of both generative and discriminative models contributes to the outstanding performance. The layered generators produce rumors by intelligently inserting controversial information in non-rumors, and force the layered discriminators to detect detailed glitches and deduce exactly which parts in the sentence are problematic. On average, in the rumor detection task, our proposed model outperforms state-of-the-art baselines on PHEME dataset by 26.85% in terms of macro-f1. The excellent performance of our model for textural sequences is also demonstrated by the gene mutation case study on which it achieves 72.69% macro-f1 score.

Introduction

In the age of rapid information propagation, rumors and misinformation spread faster than they can be verified. Traditional detection methods often rely on hand-crafted features or verified news databases, which aren't always available for "breaking" rumors. Furthermore, most AI models act as "black boxes," providing a verdict without explaining the underlying reasoning.

This research identifies a parallel between rumors and genetic mutations: both are sequences where specific "tokens" (words in a tweet or nucleotides in DNA) have been altered, creating an anomaly or "glitch." The challenge lies in creating a model robust enough to handle discrete text data while providing granular, token-level explanations for its decisions.

Methodology: A Layered GAN Framework

The proposed architecture is a layered Generative Adversarial Network (GAN) designed for sequential data. It splits tasks between specialized components:

  1. Layered Generators ($G_{where}$ and $G_{replace}$): $G_{where}$ identifies which tokens in a "clean" sequence should be changed, while $G_{replace}$ (an encoder-decoder with attention) selects the substitutes. Together, they create "intelligent" adversarial samples—rumors that mimic non-rumors—to challenge the discriminators.
  2. Layered Discriminators ($D_{classify}$ and $D_{explain}$): $D_{classify}$ categorizes the sequence (e.g., Rumor vs. Non-rumor), while $D_{explain}$ focuses on pinpointing the specific tokens that make the sequence suspicious.

To handle the non-differentiable nature of discrete text, the framework employs Reinforcement Learning (Policy Gradients), allowing the generator to receive rewards based on its ability to fool the discriminators.

Key Results and Discussion

The model was evaluated on the PHEME dataset (Twitter rumors) and genetic splice site datasets (NN269).

Performance on Rumor Detection

On the PHEME dataset, the proposed model significantly outperformed baselines such as LSTMs and VAEs. As shown in the performance comparison table, the layered GAN approach achieves superior Macro-f1 and Accuracy scores across various test scenarios.

Table of Macro-f1 and Accuracy results for various models on PHEME datasets.

The visual representation of these results highlights the consistency of the model's performance, particularly its robustness when facing "unseen" data (out-of-domain news events) compared to traditional deep learning models.

Visual comparison of model performance (Macro-f1 and Accuracy) against baselines.

Genetic Mutation Detection

In the gene mutation case study, the model achieved a 72.69% macro-f1 score. By framing genetic sequences as a language, the researchers successfully applied NLP-based anomaly detection to biological data, accurately identifying mutated tokens in DNA sequences.

Conclusion

By treating rumors and mutations as sequence anomalies, this layered GAN framework provides a unified approach to detection and explanation. The separation of classification and explanation tasks prevents "function mixture," leading to a 26.85% improvement over state-of-the-art baselines on the PHEME dataset. Future work will focus on automating hyperparameter selection and implementing Hierarchical Attention Networks to further refine the generative process.

Summary updated

Abstract

Social media have emerged as increasingly popular means and environments for information gathering and propagation. This vigorous growth of social media contributed not only to a pandemic (fast-spreading and far-reaching) of rumors and misinformation, but also to an urgent need for text-based rumor detection strategies. To speed up the detection of misinformation, traditional rumor detection methods based on hand-crafted feature selection need to be replaced by automatic artificial intelligence (AI) approaches. AI decision making systems require to provide explanations in order to assure users of their trustworthiness. Inspired by the thriving development of generative adversarial networks (GANs) on text applications, we propose a GAN-based layered model for rumor detection with explanations. To demonstrate the universality of the proposed approach, we demonstrate its benefits on a gene classification with mutation detection case study. Similarly to the rumor detection, the gene classification can also be formulated as a text-based classification problem. Unlike fake news detection that needs a previously collected verified news database, our model provides explanations in rumor detection based on tweet-level texts only without referring to a verified news database. The layered structure of both generative and discriminative models contributes to the outstanding performance. The layered generators produce rumors by intelligently inserting controversial information in non-rumors, and force the layered discriminators to detect detailed glitches and deduce exactly which parts in the sentence are problematic. On average, in the rumor detection task, our proposed model outperforms state-of-the-art baselines on PHEME dataset by [Formula: see text] in terms of macro-f1. The excellent performance of our model for textural sequences is also demonstrated by the gene mutation case study on which it achieves [Formula: see text] macro-f1 score.
BibTeX
@inproceedings{cheng2021from,
  author = {Mingxi Cheng and Yizhi Li and Shahin Nazarian and Paul Bogdan},
  title = {{From rumor to genetic mutation detection with explanations: a GAN approach}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-84993-1},
}
Figure