Availability-conditioned spatial compensation for traffic forecasting with incomplete observations

Abstract

Accurate traffic forecasting depends on reliable historical observations, yet sensor readings are often unavailable in practice. Simply filling missing entries with numerical placeholders obscures their observation state and may distort the representation of historical traffic conditions. Graph neural networks can capture spatial dependencies among traffic sensors, but relying only on the physical road graph may not provide the most effective compensation when a sensor’s own history is incomplete. We therefore propose the Availability-Conditioned Spatial Compensation Graph Convolutional Recurrent Network (ACSC-GCRN), which uses observation availability to guide both historical feature use and spatial compensation. An availability-gated encoder prevents unavailable values from contributing to the value representation while retaining the observation state. The model further complements the road graph with correlation and adaptive relations and adjusts their contributions according to the current observation condition. Experiments on METR-LA, PEMS04, and PEMS08 under standard and synthetic missingness protocols, together with an additional dataset-native invalid-reading protocol on METR-LA, show that ACSC-GCRN consistently remains among the two best-performing methods across diverse observation conditions. Ablation results support the gated encoding and multi-source spatial compensation, with availability-conditioned fusion showing greater value under severe missingness. These findings demonstrate that explicitly representing observation availability and using it to guide spatial compensation improves the robustness of traffic forecasting from incomplete histories.