Develop a discriminative base model for rare heterogeneous road users

Develop a base model for crash-severity prediction that discriminates severity among rare, heterogeneous road users sufficiently well for a valid prediction set to remain narrow and informative.

Background

The certification framework can restore coverage for heterogeneous subpopulations through conditional calibration and can transfer coverage to true severity through a declared reporting band. However, the empirical results show that the resulting composed prediction sets become nearly vacuous for severe strata such as motorcyclists and unrestrained drivers. The authors report that this saturation is not primarily caused by the reporting band: the underlying base prediction sets are already wide before band expansion.

The unresolved issue is therefore model-side rather than certification-side. A useful solution would need to improve discrimination of severity within rare and structurally heterogeneous road-user populations while preserving the validity guarantees supplied by the certification layer. The problem is included because the paper explicitly identifies it as an open problem rather than merely proposing improved modeling as desirable future work.