Effect of conflict-based neighbor selection in GATransformer

Determine whether selecting GATransformer’s neighboring vessels by predicted collision or traffic conflict, rather than by distance ranking, yields a measurable benefit from the spatial graph-attention encoder.

Background

The audited GATransformer uses distance-ranked neighbors and ten-minute sampling. Across several neighborhood sizes and occupancies, the spatial graph-attention component is inert to mildly harmful relative to a model without the spatial encoder.

The authors note that all tested arms use distance-ranked neighbor selection. Consequently, the unresolved question is whether a conflict-aware neighborhood definition could make graph attention useful under the same trajectory-prediction setting.