Dynamic Causal Explanation Based Diffusion-Variational Graph Neural Network for Spatiotemporal Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Dynamic Graph Neural Networks (GNNs) are crucial for spatiotemporal forecasting.
• Existing methods often lack explainability regarding causal relationships in dynamic graphs.
• Real-world time-series data frequently involves uncertainty and noise in dynamic graph structures.

Purpose of the Study:

• To propose a novel dynamic diffusion-variational GNN (DVGNN) for improved spatiotemporal forecasting.
• To enhance the explainability of dynamic graph construction by exploring causal relationships.
• To address the uncertainty and noise inherent in dynamic graphs.

Main Methods:

• An unsupervised generative model for dynamic graph construction using a diffusion model.
• Graph convolutional networks (GCNs) for encoding latent node embeddings and inferring dynamic link probabilities.
• A diffusion model for reconstructing causal graphs (CGs) adaptively.
• Dynamic GCN and temporal attention for future state prediction.

Main Results:

• DVGNN outperforms state-of-the-art approaches on four real-world datasets.
• Achieved outstanding Root Mean Square Error (RMSE) results and demonstrated higher robustness.
• Effectively reflects causal relationships and uncertainty in dynamic graphs, validated by F1-score and probability distribution analysis.

Conclusions:

• The proposed DVGNN model offers a robust and explainable solution for spatiotemporal forecasting.
• DVGNN successfully models causal relationships and uncertainty in dynamic graphs.
• The approach provides a significant advancement in handling complex, real-world graph-structured time-series data.