Feature Conformal Prediction Overview

• Feature conformal prediction is a framework that uses features to construct valid prediction sets with marginal guarantees, adapting to local data complexity.
• It integrates calibration strategies such as group-conditional methods, deep feature-space mapping, and neighborhood-based adjustments to enhance predictive reliability.
• Practical implementations address trade-offs between efficiency and localized coverage through methods like Mondrian CP and online adaptive algorithms.

Feature conformal prediction is the use of conformal prediction in which features enter both the construction and the interpretation of predictive sets. In a narrow sense, all conformal prediction is “feature conformal prediction”: the conformity measure $A(X,Y)$ and the prediction sets $\Gamma_{1-\alpha}(X)$ depend on features. In a broader sense, the term refers to methods that calibrate in semantic feature spaces, condition on feature-defined groups or neighborhoods, or adapt uncertainty sets to feature-local difficulty while retaining conformal guarantees. The resulting methods remain distribution-free or exchangeability-based at the level of marginal validity, but they are explicitly concerned with how coverage, set size, and reliability vary as a function of $X$ (Bellotti et al., 9 Aug 2025, Teng et al., 2022).

1. Foundations and formal setup

Conformal prediction is formulated in the standard supervised learning setting with labeled examples

$(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$

assumed exchangeable. A new example $(X_{n+1},Y_{n+1})$ is drawn from the same exchangeable distribution. A conformity measure is a function

$$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$

and it depends on features $x$. In practice, $A$ is built using a base machine learning model that has learned the relationship between $X$ and $Y$; for regression, a common choice is a normalized residual

$\Gamma_{1-\alpha}(X)$0

while for classification it is often a function of class probabilities for the pair $\Gamma_{1-\alpha}(X)$1. Inductive conformal prediction splits data into a proper training set to fit the base model and a calibration set to compute conformity scores, then defines the prediction set

$\Gamma_{1-\alpha}(X)$2

with the marginal validity guarantee

$\Gamma_{1-\alpha}(X)$3

under exchangeability (Bellotti et al., 9 Aug 2025).

The dependence on features is already present in standard classification implementations. In the random-forest-based conformalClassification package, the conformity score for class $\Gamma_{1-\alpha}(X)$4 is the vote fraction

$\Gamma_{1-\alpha}(X)$5

and prediction regions are built from classwise conformal $\Gamma_{1-\alpha}(X)$6-values. This makes feature dependence explicit: different feature vectors induce different vote distributions, different conformity scores, and therefore different prediction sets (Gauraha et al., 2018).

2. Marginal validity, conditional coverage, and feature-defined groups

The central distinction in feature conformal prediction is between marginal and conditional validity. Standard conformal prediction guarantees coverage averaged over the joint distribution of $\Gamma_{1-\alpha}(X)$7, but validity need not hold conditionally on features or subgroups. The contrast can be written as

$\Gamma_{1-\alpha}(X)$8

for marginal coverage, versus

$\Gamma_{1-\alpha}(X)$9

for group-conditional coverage, where $X$0 is a protected attribute or other feature-derived grouping. An explicit example shows a conformal predictor with overall coverage $X$1 at target $X$2, but $X$3 coverage for males and $X$4 for females. This is a feature-conditional miscalibration despite satisfactory marginal calibration (Bellotti et al., 9 Aug 2025).

This discrepancy is not an incidental defect but a theoretical boundary. Exact conditional validity is impossible in general finite-sample settings. In particular, one cannot in general require

$X$5

without degenerating into trivial predictors. This impossibility is the main reason that feature conformal prediction is organized around approximate conditional coverage, restricted conditional coverage, and group-conditional validity rather than exact pointwise guarantees (Bellotti et al., 9 Aug 2025).

A practical response is to weaken conditioning. Restricted conditional coverage asks for coverage within local neighborhoods such as metric balls $X$6,

$X$7

or within pre-specified groups. This suggests a continuum of feature-conditioning strategies: coarse partitions are statistically stable but weakly localized; fine-grained partitions are more faithful to feature space but introduce computational difficulties and often worsen efficiency through larger intervals (Bellotti et al., 9 Aug 2025).

3. Calibration in semantic feature spaces

A distinct line of work uses feature spaces internal to deep networks rather than operating directly in the output space. In “Predictive Inference with Feature Conformal Prediction,” the model is decomposed as $X$8, where $X$9 is a learned feature representation and $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$0 is a prediction head. The nonconformity score is defined by the smallest feature-space perturbation needed to reach the true label: $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$1 Calibration is performed on these feature-space distances, and a Band Estimation step maps the resulting feature-space ball back into an output-space prediction set. Under exchangeability, the resulting procedure preserves conformal coverage, and under mild assumptions it provably outperforms regular conformal prediction in average band length (Teng et al., 2022).

Fast Feature Conformal Prediction replaces the expensive nonlinear mapping step in FeatureCP by a first-order Taylor approximation. Writing $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$2, $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$3, and using the gradient of the head with respect to the feature vector, FFCP defines the nonconformity score

$(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$4

with prediction band

$(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$5

The method retains finite-sample coverage under exchangeability, performs comparably with FCP, and achieves a significant reduction in computational time by approximately $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$6 (Tang et al., 2024).

In medical image segmentation, COMPASS extends the same principle to downstream scalar metrics such as lesion area. The segmentation model is written as $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$7, and calibration is performed by perturbing intermediate features $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$8 along low-dimensional subspaces maximally sensitive to the target metric. For calibration point $(z_1,\dots,z_n)=((X_1,Y_1),\dots,(X_n,Y_n)),$9, the score is

$(X_{n+1},Y_{n+1})$0

where $(X_{n+1},Y_{n+1})$1 is the metric interval induced by feature perturbations of magnitude $(X_{n+1},Y_{n+1})$2. Under exchangeability and nestedness, COMPASS achieves valid marginal coverage; with weighted conformal prediction it can also recover target coverage under covariate shifts using importance weights estimated from learned internal features (Cheung et al., 26 Sep 2025).

4. Group-conditioned, online, and multivalid methods

A second major family of feature conformal methods does not move calibration into latent feature space, but instead calibrates explicitly with respect to feature-defined groups, neighborhoods, or time-varying relevance. Mondrian conformal prediction partitions calibration data by a taxonomy of subgroups $(X_{n+1},Y_{n+1})$3. For a new point in group $(X_{n+1},Y_{n+1})$4, only calibration scores from that group are used, yielding

$(X_{n+1},Y_{n+1})$5

under exchangeability within each group. This gives exact group-conditional validity for chosen groups, but as the taxonomy is refined, group sizes become small and efficiency deteriorates (Bellotti et al., 9 Aug 2025).

IFACM pursues approximate conditional validity by iteratively reshaping the conformity measure. Starting from a base score, it identifies feature regions with under- or over-coverage and then expands or shrinks prediction sets in those regions. In taxonomy mode the regions are pre-specified groups; in discovery mode they are learned by boosting-like splitting. Empirically, Mondrian CP and IFACM produce coverage across groups that is much closer to nominal than baseline ICP, but the price is increased inefficiency, especially in minority groups (Bellotti et al., 9 Aug 2025).

In the sequential setting, “Practical Adversarial Multivalid Conformal Prediction” develops MVP, which is computationally lightweight, uses all data for model fitting, and enforces empirical coverage simultaneously on many, possibly intersecting feature-defined groups. Its guarantees are stronger than marginal coverage in two ways: coverage is calibrated conditional on threshold buckets and conditional on membership in each user-specified subset of feature space. The resulting multivalid guarantee applies even against adversarially chosen data, and empirically MVP gives coverage in every group very close to target while maintaining smaller intervals than conservative groupwise baselines (Bastani et al., 2022).

Two recent online time-series methods combine feature representations with adaptive coverage control. AFOCP defines nonconformity in the feature space of a pre-trained neural network,

$(X_{n+1},Y_{n+1})$6

and uses an attention mechanism to adaptively weight historical observations according to their relevance to the current test point. It maintains deterministic long-term coverage under temporal dependence and distribution shifts while reducing interval size by as much as $(X_{n+1},Y_{n+1})$7 as compared to standard OCP (Zhu et al., 19 Nov 2025). FFDCI instead uses features extracted from a pre-trained deep forecasting model to fit a residual quantile predictor, then applies an online adaptive coverage control update to a per-dimension, per-horizon correction term. Its theoretical guarantee is asymptotic coverage convergence, with error bounds dependent on the feature quality of the underlying point prediction model, and experiments on 12 datasets show tighter confidence intervals while maintaining desired coverage rates (Huang et al., 13 May 2025).

5. Feature-level explanation, selection, and structured extensions

Feature conformal prediction has also been extended from interval construction to interval analysis. “Unveil Sources of Uncertainty” defines cooperative games in which conformal interval properties such as width, lower bound, and upper bound serve as value functions. These are decomposed using Harsanyi allocations, including Shapley values and proportional Shapley values, so that uncertainty itself rather than only the point prediction is attributed to individual input features. The method is model-agnostic, includes a Monte Carlo approximation with robust statistical guarantees, and provides a feature-level decomposition of quantities such as

$(X_{n+1},Y_{n+1})$8

for interval width (Idrissi et al., 19 May 2025).

A complementary direction uses conformal quantities to drive feature selection. Conformal Recursive Feature Elimination defines a feature nonconformity measure

$(X_{n+1},Y_{n+1})$9

where $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$0 is the contribution of feature $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$1 to the multiclass nonconformity score. The key identity is that the change in total nonconformity after removing feature $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$2 equals $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$3, so the algorithm recursively removes the feature with largest $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$4. The method also includes an automatic stopping criterion based on the second derivative of the mean $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$5-curve and a weighted consistency index for selected subsets (López-De-Castro et al., 2024).

The same feature-oriented logic appears in structured-output and scientific-model settings. For circular regression, projected random forests convert any linear-response model into a circular model through separate prediction of $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$6 and $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$7, then use circular-distance conformal scores and out-of-bag dynamics to avoid a separate calibration sample, producing adaptive arc-length prediction sets with approximate conformal validity in practice (F. et al., 2024). For sparse identification of nonlinear dynamics, feature conformal prediction is applied to SINDy coefficients by constructing surrogate coefficient vectors through jackknife-constrained refits and calibrating the score

$$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$8

This yields more robust uncertainty intervals for model coefficients, even under non-Gaussian noise, than standard ensemble SINDy coefficient estimates (Fasel, 15 Jul 2025).

A broader generalization is conformal loss-controlling prediction, which replaces miscoverage by an arbitrary bounded nested loss. For a nested family of set-valued predictors $$A:(x,y)\mapsto \text{“how typical is the pair }(x,y)\text{?”},$$9, CLCP guarantees

$x$0

Because the loss can encode feature- or label-dependent stakes, this framework is directly relevant to feature-conditioned reliability beyond set membership alone (Wang et al., 2023).

6. Trustworthy AI implications, limitations, and open problems

Feature conformal prediction is closely tied to trustworthy AI because it operationalizes reliable uncertainty quantification, feature-level fairness analysis, robustness to anomalies, and governance. For a given instance $x$1, conformal prediction produces an interval or label set whose size adapts to feature-local difficulty. In difficult or underrepresented regions of feature space, intervals widen; in dense or low-noise regions, intervals shrink. This makes conformal predictors useful for generalization-risk monitoring: if empirical coverage deviates from nominal, the exchangeability assumption has likely failed, and conformal $x$2-values can be aggregated into a test martingale or e-value to detect distribution change in the joint law of $x$3 (Bellotti et al., 9 Aug 2025).

The same feature dependence makes conformal prediction a diagnostic for fairness. Coverage and prediction-set size can be analyzed per subgroup defined by gender, race, age, or other protected features. Disparities such as under-coverage in minority groups are a form of biased performance, and methods such as Mondrian CP, IFACM, MVP, and feature-weighted online calibration make those disparities measurable and, to a degree, correctable. This suggests a practical role for feature conformal prediction in transparency, documentation, and risk assessment in domains such as medical AI and systems subject to regulatory review (Bellotti et al., 9 Aug 2025).

Its limits are equally clear. Exact conditional coverage is impossible in general finite-sample settings. Group-based conditional validity is guaranteed only for chosen groups, and finer partitions reduce efficiency and stability. Approximate conditional coverage and multivalidity improve feature-conditional behavior but often at the cost of larger sets, higher computational burden, or stronger algorithmic machinery. Feature-space methods additionally depend on representation quality: when the network is untrained or gradients are unstable, the benefits of FeatureCP and FFCP collapse toward vanilla conformal behavior, and tiny gradients can make normalization ill-conditioned (Tang et al., 2024).

Open directions recur across the literature. Better local coverage control in high-dimensional feature spaces, more interpretable feature-conditional diagnostics, stronger groupwise guarantees under shift, and feature-aware conformal methods for generative models and LLMs are all explicitly suggested. A plausible implication is that the field is converging on a layered view: marginal validity remains the universal backbone, while feature conditioning, semantic representations, subgroup calibration, and task-specific loss control are the main instruments for pushing conformal prediction toward finer-grained reliability without abandoning its exchangeability-based foundations (Bellotti et al., 9 Aug 2025).