Early Rumor Detection in Social Media Based on Graph Convolutional Networks

Abstract

As a popular source of news, social media provides a platform to spread and gather information, including false information due to its unrestricted fashion of information sharing. Rumor detection is a trending topic in the social networking field, and remains difficult to solve, especially challenged by massive data and large scale in social networking. Most state-of-art efforts use machine learning approaches: hand-crafted textual features from the dataset, rumor classification, and rumor prediction. However, with the availability of abundant information and unreliability in the use of hand-crafted features, models are expected to capture and leverage data features dynamically. This project demonstrates such a model implemented to work on the dynamic features of rumors. Our primary goal is to capture the influential features of rumors and learn dynamic structures built over its propagation for early rumor detection. Specifically, we use Graph Convolutional Networks (GCN) to represent the rumor propagation tree with source and response posts as graphs, and update node representations based on responses to rumors discovered over time. We use a pattern matching algorithm to detect similar sub-graph patterns generated across the graphs for efficient structure reconstruction. Over the updated graph structures is a recursive learning process through which we can accurately predict rumors. This way we deliver an efficient working model that predicts rumors earlier than existing approaches.