JSL-MGNN: A spatiotemporal graph framework for unified indoor occupant density interpolation and prediction with missing data resilience

The Internet of Things (IoT) enables real-time monitoring, but precise space management requires accurate indoor occupant density estimates. However, the sparse sensor deployment and missing data result in significant challenges in spatiotemporal data analytics due to limited data quality and inconsistencies. To address sparsity and missing data issues in real-world spatiotemporal occupant data, this study proposes a novel framework integrating spatiotemporal interpolation and prediction, leveraging shared task information to enhance resilience to data sparsity. The overall framework is centered on a model named Joint Spatiotemporal Learning Multi-Graph Neural Network (JSL-MGNN). In the evaluated comparisons, JSL-MGNN maintains strong and stable performance across both tasks. Additionally, the robustness of our model is validated under different missing data conditions. Notably, our model maintains a consistent R 2 value above 95% in the interpolation task and above 90% in the prediction task, whereas other baseline models exhibit significant fluctuations in the interpolation task. Specifically, under high missing conditions, JSL-MGNN maintains balanced performance across interpolation and prediction, with interpolation R 2 consistently above 95% and prediction R 2 above 90% in the reported settings. This framework resolves sparsity and missing data in spatiotemporal occupant data, a key advancement with notable implications for space management and occupant-centric research, enhancing analytical reliability and supporting better decision-making in occupant-centric studies.