Generalizability of leakage–capacity scaling across traffic mixes

Determine whether the exact relationship between leakage sensitivity and model capacity observed for TrAISformer on Danish AIS traffic holds under a different maritime traffic mix, and establish the causal history mechanism underlying that relationship.

Background

The paper finds that permissive vessel-sharing splits substantially reduce prediction error for the 7–8 million-parameter TrAISformer-family models, while the effect is small for the compact 0.43-million-parameter encoder. Within the tested TrAISformer family, the leakage gap also increases with model capacity, reaching a much larger effect for the 57.4-million-parameter model.

However, the capacity sweep and the causal test linking the gain to test vessels’ own history were conducted only on the Danish corpus. The authors explicitly leave unresolved whether the detailed shape of this relationship generalizes to other traffic compositions and whether the observed capacity dependence is explained by the same causal history mechanism elsewhere.

References

The fine-grained leakage-versus-capacity sweep (2\% at 0.43~M parameters, 23--25\% at 7--8~M, 67\% at 57.4~M) is measured on Danish traffic only, and the largest size on one seed; whether the relationship's exact shape, rather than its direction, holds on a different traffic mix is untested, as is the causal history mechanism behind it (Section~\ref{sec:leakage}).

— Protocol before progress: leakage-aware evaluation of AIS trajectory prediction  (2609.25827 - Raisi et al., 22 Sep 2026) in Section 6, Discussion and Limitations, subsection “Scope of the leakage audit”