EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

When Spatial Message Passing Fails to Help: Graph Attention Networks and Temporal Baselines for Dengue Forecasting

Gideon K. Gogovi

Applied Computational Intelligence and Soft Computing · 2026-01 · 2026 (1)

DOI: 10.1155/acis/2627876 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 22.07.2026. Verfasst von Gideon K. Gogovi. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Infectious disease outbreaks exhibit pronounced spatial coupling: case surges in one district frequently precede or co‐occur with surges in geographically proximate and climatically similar neighbors, yet standard time‐series models treat districts as isolated units and ignore this propagation structure. We present a spatiotemporal graph attention network (ST‐GAT) for weekly district‐level dengue fever forecasting across all Sri Lanka administrative districts, using geographic adjacency and NASA satellite‐derived climate cosine similarity as principled, complementary edge‐weight sources. A two‐layer GAT–GRU architecture is trained on 8‐week input windows to forecast one‐to‐four weeks ahead, and a systematic edge‐ablation experiment decomposes the predictive contribution of each connectivity source. Interpretable graph attention weights are aggregated over the test period to identify districts that consistently serve as spatial leading indicators of outbreak onset in their neighbors. On the held‐out test set, the district‐independent LSTM achieves the lowest aggregate RMSE (0.668, log 1+ x scale), outperforming all ST‐GAT variants, while the VAR(1) (RMSE 0.804) outperforms the ST‐GAT with combined edges (RMSE 1.000) but is in turn outperformed by the district‐independent LSTM (RMSE 0.664). Among GNN edge‐structure variants, geographic‐only edges (0.940) outperform climate‐similarity‐only (1.108) constructions. The attention analysis identifies Jaffna, Kurunegala, and Ratnapura as the leading spatial indicators, characterized by high outgoing attention in the weeks preceding outbreak onset in neighboring districts. These findings characterize the boundary conditions under which spatial graph inductive biases confer a forecasting advantage: in a single‐disease, small‐graph, distributional‐shift setting, a district‐independent LSTM outperforms both the ST‐GAT and a capacity‐matched TemporalGRU baseline, while the ST‐GAT still outperforms the TemporalGRU ( p = 0.010, Diebold–Mariano test), confirming that spatial message passing adds signal but cannot compensate for the mismatch between a joint‐district parameterization and a single‐disease surveillance context.

Zitieren

APA

Gogovi, G. K. (2026). When Spatial Message Passing Fails to Help: Graph Attention Networks and Temporal Baselines for Dengue Forecasting. Applied Computational Intelligence and Soft Computing, 2026(1). https://doi.org/10.1155/acis/2627876

Vancouver

Gogovi GK. When Spatial Message Passing Fails to Help: Graph Attention Networks and Temporal Baselines for Dengue Forecasting. Applied Computational Intelligence and Soft Computing. 2026;2026(1). doi:10.1155/acis/2627876

BibTeX
@article{Gogovi2026,
  author = {Gogovi, Gideon K.},
  title = {When Spatial Message Passing Fails to Help: Graph Attention Networks and Temporal Baselines for Dengue Forecasting},
  journal = {Applied Computational Intelligence and Soft Computing},
  year = {2026},
  volume = {2026},
  number = {1},
  doi = {10.1155/acis/2627876}
}
Kostenlos starten →