Overcoming the Data Dilemma in Traffic Flow Prediction: An MIGCN Framework Using Non-Full-Sample Floating Car Data

Accurate traffic flow prediction is essential for transitioning from reactive emergency responses to proactive traffic control. However, current forecasting approaches face a critical data dilemma: traditional fixed-point sensors lack spatial coverage, while extensive navigation Floating Car Data (FCD) inherently suffers from structural missingness due to dynamic, non-full-sample user penetration rates. To resolve this dilemma, an advanced spatial–temporal graph convolutional framework, named MIGCN, is proposed. First, to mitigate the spatial noise induced by sampling fluctuations, a mutual information (MI) algorithm is utilized to construct a global information-sharing matrix, accurately capturing non-linear statistical dependencies. Second, to dynamically compensate for the inherent missingness of FCD, critical auxiliary variables (e.g., vehicle speed and truck penetration rates) are strategically filtered and fused through contribution analysis and a self-attention mechanism. The framework further employs a refined spatiotemporal architecture, integrating spatial self-attention with graph convolution for spatial dynamics, and time-trend position encoding with pointwise attention for temporal dependencies. Empirical evaluations on real-world datasets demonstrate that MIGCN significantly outperforms baseline models, particularly in maintaining robust performance under complex operational conditions.