Research On Rumor Detection Based On Social Media Analytics And Information Propagation Structure

Nowadays,as the primary medium of information dissemination,rumors on social media platforms spread at unprecedented speed,reaching global audiences and exposing users and communities to the great risk of being misled.Therefore,it is necessary to automatically detect rumors as early as possible.Rumor detection on social media aims to predict the veracity of a given claim.Since the content of rumors itself is accompanied by uncertainty,the community response on social media with the information propagation structure can provide indicative signals for verifying the truthfulness of claims.Recently,techniques using deep learning have shown remarkable performance in rumor detection.This paper studies the rumor detection task based on social media analytics and information propagation structure,from the perspectives of modeling and encoding,respectively.We further analyze the related works with their limitations on the rumor detection task.Novel architectures are proposed concerning the modeling and encoding paradigm,which complement each other and promote the development of the rumor detection task.The main contribution is as follows:1.In terms of the modeling ways,existing methods are either limited to the strict responsive relationship or oversimplify the conversation structure to make irrelevant interactions that hinder performance improvement.In order to enhance the interaction of user opinions while mitigating the negative impact of interactions between irrelevant posts,this research proposes to represent information dissemination as an undirected interaction topology.Unlike traditional methods that ensemble models with bottom-up and top-down patterns,the undirected interaction topology allows duplex interactions between posts with responsive parent-child or sibling relationships in the conversation thread,to reinforce the interaction of user opinions and keep aware of global structural information during rumor propagation.Based on the undirected interaction topology,this work employs Graph Attention Networks for structural representation learning to predict rumor classification.Experiments show that the rumor detection model based on undirected interaction topology and graph attention network achieves superior performance on three real-world benchmark datasets for rumor detection.2.Although traditional structural encoding techniques can realize representation learning on the structure of social media information propagation,they largely ignore the wide impact of the source claims on social media conversation.To address this issue,this work proposes a Claim-guided Hierarchical Graph Attention Network to represent tweet content and information propagation structure into a latent space,which is composed of post-level claim-aware attention and event-level inferencebased attention to capture multi-level rumor-indicative signals.It enhances representation learning for responsive posts by taking the entire social context into account and attends over the posts that can semantically infer the target claim.Extensive experiments on three social media datasets show that the method further improves the rumor classification and exhibits an excellent capacity for detecting rumors in the early stage.