RumorMixer: Exploring Echo Chamber Effect and Platform Heterogeneity for Rumor Detection

Rumors have exerted detrimental effects on individuals and societies in recent years. Despite the deployment of sophisticated Graph Neural Networks (GNNs) to analyze the structure of propagation graphs in rumor detection, contemporary approaches often neglect two pivotal elements. Firstly, the structure of rumor propagation in social networks is characterized by a community-based feature, influenced by the “echo chamber effect”. By integrating these structures, models can emphasize critical information, mitigate the impact of irrelevant data, and enhance graph representation learning. Secondly, the existing models for rumor detection struggle to adjust GNN backbones to accommodate the diverse complexities introduced by social media’s platform heterogeneity. The manual design of these models is both time-consuming and labor-intensive. To overcome these challenges, this paper presents RumorMixer, a novel automated framework for rumor detection. This methodology begins by developing a Super-Sharer-Aware (SSA) chamber partitioning algorithm, crucial for identifying echo chambers within propagation graphs. Through accurate partitioning, RumorMixer effectively concentrates on the essential structures of rumor propagation and utilizes the GNN-Mixer model to create high-quality representations of these chambers. To address platform heterogeneity, RumorMixer integrates five distinct components: PE, Aggregation, Merge, Pooling, and Mixing-to establish an extensive search space. Subsequently, differentiable architecture search technology is employed to automatically tailor platform-specific architectures. The efficacy is validated through extensive experiments on real datasets from both Weibo and Twitter³(Our code is accessible at https://github.com/cgao-comp/RumorMixer.).