Normalised Sensitivity Ratio: Post-Hoc Causal Feature Identification
A new method called the Normalised Sensitivity Ratio (NSR) allows post-hoc identification of causally relied-upon features in trained models without access to training procedures. NSR is model-agnostic and operates under a structured-shift regime where environments differ primarily in spurious feature means while causal mechanisms remain stable. Under a linear structural causal model with at least three non-degenerate environments, NSR achieves exact identification. The method formalizes causal features as those inducing constant model sensitivity across environments, while spurious features track shift. Failure occurs with weak shifts, specifically O(ε⁴) collapse. The research is published on arXiv with ID 2607.25546.
Key facts
- NSR is a post-hoc, model-agnostic diagnostic for causal vs. spurious feature reliance.
- It operates under a structured-shift regime: environments differ in spurious feature means.
- Causal features induce constant sensitivity across environments; spurious features track shift.
- NSR is defined as the squared coefficient of variation of per-environment sensitivity.
- Under a linear SCM with K≥3 non-degenerate environments, NSR achieves exact identification.
- Failure occurs with weak shifts (O(ε⁴) collapse).
- The method requires no access to training procedures.
- Published on arXiv as 2607.25546.
Entities
Institutions
- arXiv