AttentionR-GCN: Incorporating Spatiotemporal Reasoning in Heterogeneous and Partially Observed Graphs

ICLR 2026 Conference Submission15056 Authors

Keywords: Heterogeneous Graph Learning

TL;DR: A graph neural network with relation-aware node and edge attention, and learnable mechanisms to handle missing values

Abstract: Urban infrastructure networks are complex systems characterized by heterogeneous nodes and edges, partial observability, and temporal dynamics, which many graph neural networks struggle to handle. We introduce AttentionR-GCN, an extension of graph attention network based on different relational types that (i) uses attention-based message aggregation to weight node and edge signals under different relation types, (ii) uses learnable embeddings to represent missing values, and (iii) incorporates a transformer encoder to model temporal dependencies. We evaluate AttentionR-GCN on two simulated water distribution networks, predicting one-step-ahead chlorine concentrations at both monitored and unmonitored nodes under varying levels of missing sensor data. Our model outperforms different baselines, especially under high data sparsity, and demonstrates superior generalization to unmonitored nodes. Our results reveal the importance of incorporating adaptive weighting of node and edge features under different relations, learnable representations for missing values, and capturing temporal dependencies to achieve more reliable predictions in partially observed infrastructure networks.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 15056