Graph Neural Networks For Relational Data Modeling And Optimization

Abstract

The digital age news systems now employ relational data models to represent news items, together with users, sources, and assertions, as interrelated objects. However, there are some challenges that might be faced when using it, such as the appearance of fake news, multimodal fake news, and Artificial Intelligence (AI) news without any context. To solve such challenges, we present the model called the Elephant Herding Optimization-Enhanced Graph Neural Network (EHO-EnGNN) that integrates Graph Neural Networks (GNNs) and Elephant Herding Optimization (EHO). Evaluation of the suggested model can be done with the help of the fake news dataset available on Kaggle, where both real and fake news examples are provided. Data preprocessing can be done via tokenization, lemmatization, and the Term Frequency–Inverse Document Frequency (TF-IDF) is used for feature extraction. Next, a graph model is built to represent the relationships between news articles, users and other entities, leading to better embeddings and classification results. The code is implemented in Python using TensorFlow, PyTorch, Compute Unified Device Architecture (CUDA), Deep Graph Library (DGL), CUDA Deep Neural Network library (cuDNN), and Scikit-learn in Jupyter Notebook and Visual Studio (VS) Code. The model's performance is promising with F1-score of 0.956, accuracy of 0.956, recall of 0.952, and precision of 0.960 in comparison with other models. In summary, the proposed method enhances the detection of fake news with enriched relational representations, graph learning and optimization.