Causal attribution can extend feature ranking toward upstream-driver analysis. Causal integrated gradients (CIG) and causal expected gradients (CEG) incorporate known causal relationships among inputs while retaining the computational approach of integrated gradients. Their design separates direct effects from effects mediated through downstream features.
The reported results are promising but bounded. In a fully controlled synthetic benchmark with a known causal model, CIG's attributions best matched the ground-truth causal effect and matched causal Shapley accuracy at a fraction of the computational cost. In interferon-β data, where only an approximate causal graph was available, CIG better prioritized transcription factors driving the measured response than DeepSHAP and integrated gradients.
Graph quality is therefore a control point, not a background assumption. In the reported synthetic and Dyngen tests, removing true edges or reversing edge direction substantially reduced recovery performance. Adding spurious edges caused a much smaller decline under those benchmark conditions. These findings do not establish how a tool will behave when an enterprise graph is incomplete, disputed, or drawn from a different domain.
Surrogate design creates a second control point. The study found systematic misallocation between direct and indirect attribution when an additive surrogate represented non-additive mechanisms. The diagnostic was strongest for interaction mechanisms, while the authors also reported that meaningful indirect-effect errors can occur in both interaction and linear settings. A high ranking can therefore reflect model and surrogate behavior as well as an upstream relationship.
For governance teams, the practical consequence is a narrower use case; use causal attribution to prioritize features, drivers, or targets for review and follow-up, while documenting the graph, surrogate assumptions, numerical settings, and evaluation scope behind each result. Treat prioritized outputs as candidates for domain review rather than definitive causal or clinical conclusions, and require validation appropriate to the intended use before relying on them for consequential decisions.