Integrating Structured Knowledge with Machine and Deep Learning for Fake News Detection

Abstract

Knowledge graphs (KGs) are key structural elements in the organization of structured information and are applicable to tasks such as semantic search, recommendation or misinformation detection, among others. In this work, we explore two such complementary paradigms characterizing KG construction and utilization:(1) classic machine learning (ML) techniques for building KGs (e.g., Support Vector Machines (SVMs) for entity recognition, Random Forests for relation extraction), and (2) deep learning (DL) approaches (e.g., BERT/RoBERTa transformers for text encoding, graph neural networks (GNNs)) for leveraging KGs in fake news detection. This survey evaluates the machine learning (ML) approaches for entity recognition, relation extraction, and knowledge representation, with a specific focus on their interpretability and scalability. We surveyed deep learning methods for fake news detection, focusing on models (BERT, RoBERTa) for contextual text analysis, graph neural networks (GNNs) and hybrid models combining KGs with sequential models. We found that while ML is a more common and reliable approach to KG construction, DL enables more complex pattern extraction for tasks such as misinformation detection.