Uncertainty-Aware Cross-Graph Learning for Conflicting Data Truth Inference in Vehicular Crowdsensing

Vehicular crowdsensing (VCS) has become an emerging paradigm that exploits various sensor-equipped vehicles to accomplish challenging sensing tasks. However, VCS usually provides conflicting information since source discrepancies, prompting truth inference as an essential tool for discovering accurate answers. Unfortunately, although numerous graph-based truth inference schemes have been proposed and achieved strong performance, they still suffer from two limitations: the inability to capture complete information between multiple data sources, and the lack of credibility assessment of the inference. This paper presents an uncertainty-aware cross-graph learning (UCGL) for conflicting data truth inference in VCS. Specifically, we propose a hierarchical fine-grained multi-data source fusion framework, including intra-view feature integration and information propagation across views. Wherein, the designed feature-enhanced graph attention encoder integrates nodes with varying importance to generate latent feature representations for each view. Moreover, we develop a cross-view adaptive fusion module that exploits a learnable soft subgraph reconstruction mechanism to capture consistent and complementary information between heterogeneity views. Based on evidence-based deep learning, a model confidence assessment scheme is introduced to guarantee the trustworthiness of the inference results. Ultimately, we conduct comparison experiments with eight state-of-the-art programs, and the performance evaluation demonstrates that our UCGL algorithm is superior and practical in VCS.