CAA-TDI: Confidence-Aware Attention for Trustworthy Data Inference in Sparse Mobile Crowdsensing

In sparse mobile crowdsensing (SMCS) systems, existing data inference models rely on reliable sensing data to estimate missing data. However, in practice, untrustworthy mobile workers intentionally submit low-quality or even false data, posing a significant threat to the accuracy of data inference. To address this challenge, we propose the Confidence-Aware Attention-based Trustworthy data inference (CAA-TDI) framework to achieve high-quality inference in the presence of false data, which contains the Spatiotemporal Confidence Learning Network (ST-CLN) and the Spatiotemporal Confidence-enhanced Attention Network (ST-CAN). Specifically, the ST-CLN first employs multiple spatiotemporal encoders to quantify data reliability as learnable confidence scores from spatial and temporal perspectives, enabling accurate noise estimation even in highly sparse datasets. Then, the ST-CAN integrates confidence scores into the multi-head self-attention mechanism, which enhances the spatiotemporal pattern learning from low-noise data, enabling high-quality inferred data. Finally, we take the theoretical analysis and conduct extensive experiments on real-world datasets under both single-task and multi-task data inference situations to validate the effectiveness of the proposed CAA-TDI framework in improving data inference quality.