CaST: Causal Discovery via Spatio-Temporal Graphs in Disaster Tweets

TL;DR

CaST addresses the challenge of discovering cause–effect relationships in disaster events from social media by unifying semantic, spatial, and temporal signals in a spatio-temporal event graph. It encodes events with CrisisTransformer-derived semantics, location and time features, and learns causal links with a two-layer multi-head Graph Attention Network, optimized via focal loss to handle class imbalance. On a 167K tweet Hurricane Harvey dataset annotated with MAVEN-ERE-based intra-tweet causal relations, CaST outperforms eight baselines on accuracy, precision, recall, F1, and AUC, with ablations confirming the value of spatio-temporal signals for stability. This approach improves robustness and interpretability of disaster propagation modeling from social media, offering a practical pathway for real-time situational awareness and decision support.

Abstract

Understanding causality between real-world events from social media is essential for situational awareness, yet existing causal discovery methods often overlook the interplay between semantic, spatial, and temporal contexts. We propose CaST: Causal Discovery via Spatio-Temporal Graphs, a unified framework for causal discovery in disaster domain that integrates semantic similarity and spatio-temporal proximity using Large Language Models (LLMs) pretrained on disaster datasets. CaST constructs an event graph for each window of tweets. Each event extracted from tweets is represented as a node embedding enriched with its contextual semantics, geographic coordinates, and temporal features. These event nodes are then connected to form a spatio-temporal event graph, which is processed using a multi-head Graph Attention Network (GAT) \cite{gat} to learn directed causal relationships. We construct an in-house dataset of approximately 167K disaster-related tweets collected during Hurricane Harvey and annotated following the MAVEN-ERE schema. Experimental results show that CaST achieves superior performance over both traditional and state-of-the-art methods. Ablation studies further confirm that incorporating spatial and temporal signals substantially improves both recall and stability during training. Overall, CaST demonstrates that integrating spatio-temporal reasoning into event graphs enables more robust and interpretable causal discovery in disaster-related social media text.

Figures (5)

• Figure 1: Example of causal relationships in disaster tweets with spatio-temporal context. Two tweets from the same location (Houston) and close temporal proximity (2 hours apart) contain causally related events. The green dashed arrow illustrates the causal inference between "Heavy rain" and "Floodwaters" across tweets, enabled by shared spatio-temporal features. Gray dashed lines show how spatial and temporal context enriches event representations to improve causal discovery.
• Figure 2: Overview of the proposed CaST framework for causal discovery in disaster tweets. The Feature Extraction module extracts events along with triggers and contextual attributes (spatial and temporal attributes), encodes them as heterogeneous nodes. A spatio-temporal graph is then constructed where black edges denote contextual link (semantic, spatial, or temporal relation), and red edges indicate predicted causal links in the output. The Graph Attention Network propagates information across these edges to learn context-aware representations that enable accurate causal link classification.
• Figure 3: JSON schema of the annotated disaster tweet dataset used in CaST. Each tweet includes the identification number, detected events, causal relations, spatial and temporal metadata, and token-level causal role masks.
• Figure 4: ROC–AUC comparison of CaST variants during training.
• Figure 5: Training and validation loss curves for CaST and its ablated variants.