Development of an Unsupervised Learning-Based Fake News Generation Framework and Study the Effectiveness of Dataset Quality Evaluation Metrics

With the advent of the internet and social media, fake news has become easily generated and disseminated, causing severe societal issues. The development of robust fake news detection models to mitigate the damages incurred from fake news necessitates a substantial volume of high-quality fake news data. However, there exists a significant shortage of publicly available datasets that embody such quality. In this study, we propose an unsupervised learning-based framework for generating fake news, solely leveraging real news data. This proposed framework is structured into four key phases: similar news exploration, fake news title generation, labeling error filtering, and data quality verification. Moreover, we introduce a novel set of evaluation metrics to assess the quality of the generated fake news data in terms of reliability, diversity, and complexity. Through extensive experiments, we have validated the effectiveness of the proposed framework and evaluation metrics, paving the way for more nuanced approaches to the creation and analysis of fake news datasets.