TrendGNN: Towards Understanding of Epidemics, Beliefs, and Behaviors

TL;DR

TrendGNN tackles the interpretability gap in epidemic forecasting by modeling interdependent signals (epidemic, behavior, and beliefs) with trend-similarity graphs and GraphSAGE. The framework constructs state-wise, signal-level graphs using DTW+S, explores multiple graph-construction strategies, and evaluates predictions over a four-week horizon with rolling-window training on CTIS data. It demonstrates that semantically meaningful graphs improve mid- to long-term accuracy and provides edge- and signal-level explanations via CF-GNNExplainer, linking outcomes to plausible drivers like mortality, risk perceptions, and trust in institutions. The work offers a principled path toward interpretable, behavior-inclusive epidemic modeling and informs public health decision-making and simulation development.

Abstract

Epidemic outcomes have a complex interplay with human behavior and beliefs. Most of the forecasting literature has focused on the task of predicting epidemic signals using simple mechanistic models or black-box models, such as deep transformers, that ingest all available signals without offering interpretability. However, to better understand the mechanisms and predict the impact of interventions, we need the ability to forecast signals associated with beliefs and behaviors in an interpretable manner. In this work, we propose a graph-based forecasting framework that first constructs a graph of interrelated signals based on trend similarity, and then applies graph neural networks (GNNs) for prediction. This approach enables interpretable analysis by revealing which signals are more predictable and which relationships contribute most to forecasting accuracy. We believe our method provides early steps towards a framework for interpretable modeling in domains with multiple potentially interdependent signals, with implications for building future simulation models that integrate behavior, beliefs, and observations.

Figures (4)

• Figure 1: Overview of the TrendGNN pipeline. The dark gray block denotes a block-diagonal matrix, where each light gray sub-block on the diagonal is an $N_{\text{signal}} \times N_{\text{signal}}$ similarity matrix for one state (computed using DTW+S). Off-diagonal regions are set to zero, and stacking across $N_{\text{state}}$ yields a final square matrix of size $(N_{\text{signal}} \times N_{\text{state}})^2$. The Input, Output, and Target in the diagram each have size $(N_{\text{state}} \times N_{\text{signal}}) \times \text{window}$, where the window length is four weeks as shown.
• Figure 2: 3D visualization of signal graphs across four categories of indicators. Each subplot corresponds to one category. Nodes are colored by category, with the center node (focus signal) highlighted. Thick edges denote important connections.
• Figure 3: Mean absolute error (MAE) distributions of different models across 1–4 week-ahead forecasting tasks.
• Figure 4: Distribution of relative improvements in MAE for different models over the baseline across 1–4 week-ahead forecasting tasks. The x-axis shows relative improvement, the y-axis shows signal categories, and colors denote models. Positive values indicate gains over the baseline, negative values indicate degradation.