ConformaDecompose explains why a conformal interval is wide: localizing calibration support per test instance shrinks width by up to about 1%
Synopsis
ConformaDecompose turns the single global calibration threshold of conformal prediction into a per-instance localization trajectory: starting from the full calibration set, it progressively downweights calibration clusters by model-aware similarity and accepts only steps that do not increase interval width, splitting width into a part removable by localization and a part that persists. Across benchmark and real-world data, the absolute reducible width aligns strongly with a local model-instability proxy while the reducible ratio varies by task; against local and weighted conformal baselines, interval width contracts by up to about 1% while coverage stays near nominal.
Interpretation
A conformal interval width stops being just a number and becomes a per-instance question: how much of that width comes from calibration regions behaviorally irrelevant to the test instance, and how much survives localization. Conformal prediction previously aggregated residuals across the whole input space into one global calibration threshold; normalized conformal prediction and conformalized quantile regression make widths instance-dependent, but uncertainty remains a single-shot outcome that does not show how calibration support contributes to width. On benchmark datasets including power plant, Parkinson telemonitoring, airfoil self-noise and abalone, plus a real Craigslist used-car dataset, the absolute reducible width correlates with the local model-instability proxy typically above 0.9, while the reducible ratio ranges from negative to strongly positive.
The procedure keeps every calibration sample and changes only its weight, holding clusters behaviorally close to the test instance at unit weight and pressing the rest toward a small constant. Each calibration point is mapped into a model-aware space that jointly encodes input location, model prediction and predictive scale, clustered there and ranked by relevance to the test instance, rather than discarding samples or recalibrating from a strictly local subset. Localization advances by a greedy accept-revert rule that accepts only steps where interval width does not increase, yielding a non-increasing width trajectory; experiments run on all test instances with Random Forest and Ridge Regression base models at clustering granularities K=5 and K=8.
Against local and weighted conformal baselines, the decomposition gives an interpretable composition of width while keeping coverage near nominal. Apparent interval narrowing by some baselines on the Parkinson data largely comes from aggressive shrinking that removes aleatoric noise rather than addressing reducible epistemic uncertainty. Across datasets, interval width shrinks by up to about 1.0% while empirical coverage stays near nominal.
Perspective
The decomposition targets regression tasks and is demonstrated and evaluated on benchmark datasets including power plant, Parkinson telemonitoring, airfoil self-noise and abalone, plus Craigslist used-car pricing, with Random Forest and Ridge Regression base models at clustering granularities K=5 and K=8; localized intervals serve explanatory analysis only and carry no new conformal guarantee.
Ratio alignment varies from negative to strongly positive across datasets and granularities, which the authors attribute to dilution by irreducible components and localization resolution; reducibility patterns depend on model-derived uncertainty proxies and clustering-based localization, all alignments are associative rather than causal, and the method does not capture uncertainty from model misspecification or limited capacity.
