BiGGAT: Bimodal Gated Graph Attention

• The paper introduces BiGGAT, which combines cluster-conditioned bimodal embedding with GRU-based gated graph attention to predict 3-class outage durations.
• It captures high-order spatial dependencies and heterogeneity in county-level disaster data by leveraging a county adjacency graph and unsupervised clustering.
• Empirical results demonstrate improved balanced accuracy and macro F1 scores over baselines, particularly for medium and long outage classes.

BiGGAT, short for Bimodal Gated Graph Attention Network, is a graph neural network architecture proposed for predicting the duration of large-scale power outages induced by natural disasters, specifically hurricanes, at the county level. The task is formulated as a node-level graph classification problem on a county adjacency graph, with county features as node attributes and outage-duration categories as labels. In the reported formulation, the prediction target is not continuous duration but a 3-class duration label—short ($<2$ days), medium ($2\text{–}6$ days), and long ($>6$ days)—chosen to align with restoration guidance and operational assessment practice. BiGGAT combines a two-cluster initial embedding scheme with graph attention and GRU-based recurrent message propagation, and is evaluated in an inductive learning regime over six major hurricanes in the Southeastern United States (Duan et al., 14 Mar 2026).

1. Task formulation and operational setting

The underlying prediction problem is defined as

$\boldsymbol{Y} = M(\boldsymbol{X}, G)$

where $G=\{V,E,\boldsymbol{A}\}$ is the graph of affected counties, $\boldsymbol{X}$ denotes county-level features, and $\boldsymbol{Y}$ denotes outage-duration labels. Each node corresponds to an affected county, and edges connect neighboring counties. This casts outage-duration estimation as a graph-structured supervised learning problem rather than a conventional tabular regression problem.

The paper emphasizes four practical difficulties. First, outage duration exhibits high-order spatial dependency: correlation is not limited to immediately adjacent counties. Second, there is marked spatial heterogeneity: counties near the storm track may experience outages lasting days, whereas counties farther away may recover in hours. Third, severe outage events are rare, so the number of large-scale events available for training is only moderate. Fourth, hurricanes and the regions they affect are heterogeneous, so the model must generalize across event-specific and region-specific impact patterns rather than fit a single stationary spatial process (Duan et al., 14 Mar 2026).

These choices clarify two points that are often misunderstood. BiGGAT is not framed as a fine-grained duration regressor; it is explicitly a 3-class classifier. Likewise, its deployment objective is not within-event interpolation alone, but generalization to unseen hurricane events.

2. Data representation, graph construction, and spatial dependence

The empirical study uses field data from six major Atlantic hurricanes in the Southeastern United States: Florence, Irma, Laura, Michael, Sally, and Zeta. Together, these events affected 557 counties and about 15 million customers. Each county node has a feature vector

$\boldsymbol{x}_v \in \mathbb{R}^{11}$

comprising: the maximum number of affected customers, county population, county area, four social vulnerability variables, and four hurricane wind-swath variables. The data sources are explicitly specified: outage counts from PowerOutage.com / PowerOutages.US, weather and wind swath from NHC best-track data (HURDAT2), population, area, and adjacency from the U.S. Census Bureau, and socioeconomic vulnerability from the CDC/ATSDR Social Vulnerability Index (SVI).

The spatial graph is written as

$G=\{V,E,\boldsymbol{A}\}$

with $\boldsymbol{A}_{ij}=1$ when counties $2\text{–}6$0 and $2\text{–}6$1 are neighbors and $2\text{–}6$2 otherwise. The construction is therefore geographic and county-adjacency based, rather than utility-topology based.

A central empirical motivation for BiGGAT is the presence of higher-order neighborhood dependence. The paper adapts Moran’s I to define nonredundant $2\text{–}6$3-hop neighborhoods and computes global $2\text{–}6$4-hop Moran’s I for peak outage counts. Significant positive spatial correlation is reported at least up to the second-order neighbor level, with the highest significant order varying by event: Florence $2\text{–}6$5, Irma $2\text{–}6$6, Laura $2\text{–}6$7, Michael $2\text{–}6$8, Sally $2\text{–}6$9, and Zeta $>6$0 (Duan et al., 14 Mar 2026). This empirical finding is the paper’s principal justification for going beyond one-hop message passing.

The broader significance is methodological. The county adjacency graph is not used merely as a convenient regularizer; it is intended to encode measurable spatial structure in outage behavior. This suggests that the model’s design is tightly coupled to the observed geography of disaster impact rather than imported generically from graph learning.

3. Bimodal embedding and the meaning of “bimodal”

BiGGAT contains two main components: Bimodal Embedding and a Gated Graph Attention mechanism. The first of these gives the model its name. For each node $>6$1, the feature vector $>6$2 is mapped to an initial message $>6$3 through cluster-specific linear embeddings. Counties are partitioned by K-means on wind-swath and duration-label structure, producing two clusters intended to reflect distinct spatial impact regimes. If $>6$4 denotes the cluster label, then

$>6$5

Each cluster therefore has its own learnable weight matrix $>6$6.

The term “bimodal” has a specialized meaning in this model. It does not refer to two input modalities such as text and images. Instead, it refers to a two-cluster embedding structure derived from unsupervised clustering of wind-swath and outage-duration patterns (Duan et al., 14 Mar 2026). This distinction matters because “bimodal” in other machine learning literatures usually denotes multimodal data fusion, which is not what is being done here.

The paper also notes that the number of clusters can be varied to reflect event complexity, although the illustrated instantiation uses two clusters. This suggests that the bimodal design is a concrete architectural choice within a broader cluster-conditioned embedding idea, but the reported BiGGAT formulation is explicitly two-cluster.

Conceptually, the bimodal embedding is the mechanism used to encode spatial heterogeneity before graph propagation begins. Counties assigned to different impact regimes are not forced through a single shared input projection. The reported improvement from GAT to BiGAT is presented as evidence that this heterogeneity modeling is useful.

4. Gated graph attention mechanism and learning procedure

After initialization, BiGGAT performs recurrent message passing in which attention-based aggregation and GRU-based state update are applied sequentially. At iteration $>6$7,

$>6$8

and

$>6$9

where $\boldsymbol{Y} = M(\boldsymbol{X}, G)$0 is the neighborhood of node $\boldsymbol{Y} = M(\boldsymbol{X}, G)$1. The attention component determines how much each neighbor contributes, while the GRU controls how newly aggregated information is fused with the previous message state.

The reported intuition is explicit. The GAT component provides local adaptivity, learning which neighboring counties are more relevant to a county’s outage duration. The GRU provides iterative memory, enabling the model to accumulate information across multiple propagation steps. The paper identifies this combination as the model’s core novelty and argues that it helps retain useful information while avoiding uncontrolled oversmoothing (Duan et al., 14 Mar 2026).

The end-to-end pipeline is summarized as follows: build the county graph; cluster counties into bimodal groups; embed node features into initial messages; apply graph attention; update messages with GRU; repeat for multiple recurrent iterations; and read out class probabilities. The node-level output $\boldsymbol{Y} = M(\boldsymbol{X}, G)$2 is produced by a linear readout layer, and training uses cross-entropy loss with the ADAM optimizer. The task is explicitly 3-class classification, with evaluation metrics of Classification Accuracy, Macro F1 Score, and Balanced Accuracy. Balanced accuracy is highlighted because the classes are imbalanced, with many more short-duration examples than medium- or long-duration examples.

The experimental regime is inductive rather than transductive. Training is performed on five hurricanes, testing on the held-out sixth hurricane, and the procedure is repeated until each hurricane has served once as test data. For test counties, the cluster label $\boldsymbol{Y} = M(\boldsymbol{X}, G)$3 is inferred from the input features. This setup is operationally important because the intended deployment scenario is prediction for a future unseen event, not prediction within a fixed graph already observed during training.

5. Empirical performance and generalization behavior

BiGGAT is compared against four baselines: the non-graph models XGBoost (XGB) and Random Forest (RF), and the graph models GAT and BiGAT. Averaged across the six held-out test events, the reported results are: XGB with 79.7% Accuracy, 67.8% Balanced Acc, and 0.657 Macro F1; RF with 81.5%, 69.4%, and 0.682; GAT with 83.7%, 68.1%, and 0.678; BiGAT with 85.1%, 70.0%, and 0.686; and BiGGAT with 85.9%, 78.3%, and 0.762 (Duan et al., 14 Mar 2026).

The comparative pattern is as important as the absolute values. The graph-based models outperform the tree-based baselines, which the paper interprets as evidence that spatial structure is informative for outage-duration prediction. The transition from GAT to BiGAT supports the value of bimodal embedding for modeling heterogeneity. The transition from BiGAT to BiGGAT supports the value of GRU-based gated propagation, with the gains described as particularly strong for balanced accuracy and macro F1, especially on the medium and long outage classes. Since those classes are the minority classes, the reported improvement is not merely aggregate but class-sensitive.

The generalization analysis further subdivides test counties into “test with overlap” and “absolute disjointed” groups. The absolute-disjointed group contains 265 counties and is described as the hardest setting because it has no geographic or structural connection to the training set. Even in this group, the model achieves 82.4% test accuracy, and those counties account for only 27.3% of all test errors (Duan et al., 14 Mar 2026). This is presented as evidence that BiGGAT generalizes beyond counties that are structurally close to training examples.

A plausible implication is that the model is learning transferable graph-conditional patterns rather than relying solely on memorized regional templates. The inductive protocol is central to that interpretation.

6. Interpretation, related gated graph attention models, and limitations

The paper does not report a detailed feature-importance analysis in the excerpt, but it does provide architecture-level evidence through comparative results. The GAT vs. BiGAT comparison is interpreted as showing that bimodal embedding captures heterogeneous spatial regimes. The BiGAT vs. BiGGAT comparison is interpreted as showing that the GRU plus attention mechanism captures higher-order dependency and improves class-balanced performance. In this sense, the model’s main interpretability claim is structural rather than attributional: different architectural components correspond to different hypothesized properties of the outage process.

BiGGAT also belongs to a broader family of gated graph attention models. A related example is the sentence-matching model sometimes summarized as Match-Graph, which uses a Gated Graph Attention Network (G-GAT) to combine attention with relation-aware message gating over a graph built from intra-sentence and cross-sentence edges (Cui et al., 2020). In that model, attention assigns importance scores to neighbors, while a learned gate modulates which dimensions of the message pass through. This suggests that “gated graph attention” is not a single canonical mechanism but a family of architectures that combine attention-based neighbor weighting with gating-based control of message propagation. BiGGAT instantiates that family in a different way: its gating is implemented through a GRU over recurrent graph states, and its “bimodal” component is a cluster-conditioned node embedding, not a relation-type gate.

Several future directions are explicitly proposed for the outage-prediction setting. The authors suggest extending the framework to other disaster types, including wildfires, winter storms, and tornadoes; using finer spatial resolution, such as utility-owned private outage data rather than county-level aggregates; and connecting predicted outage impacts to resilience and vulnerability analysis for infrastructure planning and assessment (Duan et al., 14 Mar 2026). A broader limitation implied in the summary is that the present formulation is county-level rather than utility-network-level, and that cross-event dependencies are not modeled explicitly. That implication is not stated as a formal theorem or ablation result, but it is consistent with the reported setup.

Taken together, these points position BiGGAT as a specialized graph neural architecture for a geographically structured, event-sparse, class-imbalanced prediction problem. Its defining contribution is the combination of cluster-conditioned heterogeneity modeling with attention-guided recurrent graph propagation under an explicitly inductive cross-event evaluation protocol.