A Hybrid Data Mining And Deep Learning Framework For Automated Fake News Detection On Social Media Platforms

Abstract

The virulent progress of the social media has greatly increased the proliferation of fake information to pose a serious challenge to the credibility of people, democracy, and social stability. In current methods of fake news detection that solely rely on traditional machine learning models or deep learning models, there are serious limitations such as poor interpretability, sensitivity to data skew, inability to generalise to other systems, and resistance to adversarial examples of fake news. In order to overcome these limitations, this paper lays out a solid hybrid system of data mining and deep learning as an efficient method of fake news automatic detection on social media. The suggested framework is a close combination of feature-based data mining algorithms and context sensitive deep semantic representation learning that aims to represent explicit features of credibility, as well as latent semantic regularities. To be more specific, the model involves and utilises statistics data mining techniques to extract multi-dimensional linguistic characteristics and user credibility characteristics as well as propagation-based characteristics, and additionally, relies on the platforms of transformer to learn deep contextual representations of the news material. Attention-based feature fusion algorithm is also used to successful focus and weight hand crafted credibility features into deep semantic embeddings, enhancing robustness and interpretability. A neural decision layer is then run on the hybrid representation to classify the representation with an optimism towards highly imbalanced social media data. Massive experimental tests performed on various benchmark data sets prove that the suggested framework can achieve significantly greater accuracy, precision, recall, F1-score and is less susceptible to noisy and adversarial samples in comparison with the state of art standalone machine learning and deep learning models. Moreover, in the course of qualitative analysis, one can see that the inclusion of propagation dynamics and linguistic credibility cues can dramatically increase the capabilities of the model to identify the patterns of subtle misinformation, as well as have the meaningful explanatory explanations. The findings show that the suggested hybrid system presents a flexible, explainable, and trustworthy tool to find fake news in reality, which is appropriate and adaptable to be implemented in the contemporary social media surveillance and content regulation platforms.