Publikasjonsdetaljer
- Journal: Advances in Neural Information Processing Systems, 2026
-
Internasjonale standardnumre:
- Trykt: 1049-5258
-
Lenke:
- ARKIV: hdl.handle.net/11250/5599497
Modern probabilistic machine learning models increasingly produce multivariate
outputs with complex dependence structures, from multi-step time-series forecasts
to sample path predictions. Understanding which input features drive the predictive
uncertainty is important for risk-aware decisions, model diagnostics, and deciding
whether the uncertainty should be mitigated or hedged against. This attribution
problem requires a choice of how dependencies between output components are
treated. Existing approaches reduce the output to a scalar through aggregation or
projection before attribution, thereby obscuring whether features affect marginal
uncertainty, dependence structure, or both, while component-wise analyses can
miss dependence effects entirely. We close this gap by introducing a hierarchy
of three entropy-based Shapley games, ranging from per-component marginal
through sequential (for ordered outputs) to fully joint entropy, that make this
output-side choice explicit. The hierarchy isolates a cross-component attribution
that captures how each feature shifts the dependence between output components,
a quantity invisible to per-component marginal attribution. We establish a chain-
rule decomposition of the joint attribution, characterize the cross-component term
through conditional total correlation, and provide both closed-form and sample-
based estimators. Finally, we evaluate the estimators against analytical, semi-
synthetic, and model-based references, demonstrate the hierarchy on real-world
distributional regression, and reveal differences in cross-component attribution
across probabilistic models, including a zero-shot time-series foundation model.