DISCO: Comprehensive and Explainable Disinformation Detection

Heterogeneous graph neural networks with denoising for graph embeddings

“…By building and analysing knowledge graphs over news content, knowledge-driven methods can provide explanations about why a news article is classified as fake news, thus enabling result interpretability. GET [10] and DISCO [58] offer result interpretability by detecting the misleading words, as mentioned above. With the assistance of the attention mechanism, word entities in the knowledge graphs assigned with higher weights are considered candidates for misleading words.…”

“…where y is the ground-truth news veracity, ŷ is the prediction result, λ reg is the regularisation coefficient, and ||Θ|| 2 is the L 2 norm of the model parameters, i.e., λ reg ||Θ|| 2 is the L 2 regularisation. GCN-text [8], GET [10], DISCO [58] all utilise this standard cross-entropy loss, i.e., setting λ reg to 0 in Equation 1, while FinerFACT [12] uses a non-zero value for λ reg . We discuss these methods below.…”

“…With the assistance of the attention mechanism, GET is able to tell which words provide fake information (i.e., the misleading words). In a similar fashion, DISCO [58] also constructs a graph to model word relationships in news content. It only extracts knowledge entities from source news, and hence only one knowledge graph is constructed for each news article.…”

“…This is based on the self-correction ability of propagation, i.e., fake news usually receives more objections and alternative opinions than real news. Other studies [10], [58] define interpretability of news content based on capturing misleading words.…”

Natural and Artificial Dynamics in Graphs: Concept, Progress, and Future

“…Those tasks can be directly adapted to solve many highimpact problems in real-world settings. For example, through learning the graph distribution and adding specific domain knowledge constraints, graph generators could contribute to molecule generation and drug discovery (Liu M. et al, 2022;Luo and Ji, 2022); With modeling picture pixels as nodes, graph partitioning algorithms could achieve effective image segmentation at scale (Bianchi et al, 2020); By modeling the information dissemination graph over news articles, readers, and publishers (Nguyen et al, 2020) or modeling the suspicious articles into word graphs (Fu et al, 2022a), node and graph classification tasks can help detect fake news in the real world.…”