Combining Topological Signature with Text Embeddings: Multi-Modal Approach to Fake News Detection

In recent decades, the online sphere's influence has grown significantly. Consequently, the proliferation of online fake content poses a serious challenge to modern society, raising significant concerns about democracy and public health. The escalating need for large-scale fact-checking has spurred the rapid development of automated solutions leveraging technologies like Natural Language Processing and Machine Learning to reduce human effort. This study delves into a novel approach to fake news detection by examining the topology of news articles. Topological Data Analysis is an emerging discipline supported by robust mathematical methods. It employs topology and geometry to capture the essence of complex and multidimensional data through topological signatures. Specifically, we investigate the effectiveness of topological signatures extracted using Persistence Homology, a fundamental technique in Topological Data Analysis, for training Machine Learning models for fake news detection. We evaluate whether their standalone use or addition to existing state-of-the-art text representation methods offers further insights compared to baseline systems. Through empirical evaluation using two real-world datasets, we found that while topological signatures are not as effective as text embedding techniques when used alone for classification, they can enhance the performance of fake news detection models when combined with text embeddings.