Transductive Conformal Prediction

• Transductive conformal prediction is a method that recalculates predictive p-values by augmenting the training data with each test point to guarantee finite-sample marginal validity under exchangeability.
• It employs class-specific nonconformity scores, using techniques like random forests, but is computationally intensive due to the need for model refitting for every candidate label.
• The approach offers robust prediction sets compared to inductive methods, balancing statistical efficiency with practical challenges in computational cost and score design.

Searching arXiv for the cited transductive conformal prediction sources to ground the article in published work. Transductive conformal prediction is the classical, full conformal formulation in which the predictive assessment for a test object is recomputed separately for each hypothesized label by temporarily augmenting the observed sample with that labeled test point and measuring how well the resulting sample conforms to exchangeability (Sesia et al., 25 Mar 2026). In supervised learning notation, observations are typically written as $Z_i=(X_i,Y_i)\in\mathcal X\times\mathcal Y$, with $X_{n+1}$ observed and $Y_{n+1}$ unknown, and the central object is the augmented dataset $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$ for a candidate label $y$ (Sesia et al., 25 Mar 2026). The method produces conformal $p$-values and prediction sets that satisfy finite-sample, distribution-free marginal coverage under exchangeability, but it is computationally demanding because the score function may depend on the full hypothetical dataset $D(y)$, so models may need to be refit for each test case and candidate label (Sesia et al., 25 Mar 2026).

1. Formal definition and foundational assumptions

Transductive conformal prediction is the form of conformal prediction that many authors call full conformal (Sesia et al., 25 Mar 2026). It is called transductive because “the model is refit separately for each test case and hypothesized label, incorporating it into the training data” (Sesia et al., 25 Mar 2026). In the classification formulation implemented in the conformalClassification package, the object space is $\mathcal{X}\subset\mathbb{R}^p$, the label space is $\mathcal{Y}=\{1,2,\dots,l\}$, and the observation space is $\mathcal{Z}=\mathcal{X}\times\mathcal{Y}$ (Gauraha et al., 2018). Training data are $X_{n+1}$0, $X_{n+1}$1, and the goal is to predict the unknown label of a new object $X_{n+1}$2 (Gauraha et al., 2018).

The basic stochastic assumption is exchangeability. In the general formulation, $X_{n+1}$3 are exchangeable if their joint distribution is invariant under permutations (Sesia et al., 25 Mar 2026). No parametric model is required; exchangeability is the central assumption underpinning the conformal validity guarantee (Sesia et al., 25 Mar 2026). The paper on conformalClassification likewise states that exchangeability, a weaker condition than i.i.d., is the standard assumption under which conformal prediction guarantees validity, meaning that the coverage probability of the prediction sets is at least the nominal level (Gauraha et al., 2018).

A nonconformity score in the general framework is a function $X_{n+1}$4, where larger values indicate that an observation is more atypical relative to the reference dataset (Sesia et al., 25 Mar 2026). The conformalClassification exposition uses a nonconformity measure $X_{n+1}$5, but notes that in practice it works mainly with a conformity measure, where larger values mean that a labeled example is more typical (Gauraha et al., 2018). This suggests that transductive conformal prediction should be understood more generally as a ranking procedure over hypothetical augmented samples rather than as a commitment to one particular sign convention.

For each candidate label $X_{n+1}$6, the hypothetical full dataset is $X_{n+1}$7 (Sesia et al., 25 Mar 2026). Scores are computed for every point relative to this same dataset, and the conformal $X_{n+1}$8-function is then defined by ranking the test score among the scores of the observed data (Sesia et al., 25 Mar 2026). In the generic formulation,

$X_{n+1}$9

The corresponding prediction set is

$Y_{n+1}$0

(Sesia et al., 25 Mar 2026). In the classification-specific presentation of TCP, the same logic appears in the class-conditional formula

$Y_{n+1}$1

(Gauraha et al., 2018).

2. Validity, coverage, and interpretation of guarantees

The principal theoretical property of transductive conformal prediction is finite-sample marginal validity under exchangeability. In the general theory, the conformal $Y_{n+1}$2-value at the true label is super-uniform: $Y_{n+1}$3 (Sesia et al., 25 Mar 2026). Equivalently, the prediction set satisfies

$Y_{n+1}$4

(Sesia et al., 25 Mar 2026). Under almost-surely distinct scores, the coverage is sandwiched by

$Y_{n+1}$5

(Sesia et al., 25 Mar 2026). This is finite-sample, exact up to $Y_{n+1}$6, and distribution-free under exchangeability (Sesia et al., 25 Mar 2026).

The classification formulation in conformalClassification states the analogous guarantee as

$Y_{n+1}$7

under exchangeability (Gauraha et al., 2018). The package-level paper further distinguishes three types of output at a given significance level: empty prediction sets, singleton sets, and multiple-label sets (Gauraha et al., 2018). This operational distinction is central in applications because coverage alone does not determine whether the prediction set is informative.

A recurrent interpretational point in the recent literature is that this guarantee is marginal rather than feature-conditional. The 2026 statistical overview explicitly states that the probability in the coverage guarantee is taken jointly over training data and the test point, and that exact conditional coverage given $Y_{n+1}$8 is generally impossible to achieve nontrivially without stronger assumptions (Sesia et al., 25 Mar 2026). This is important because transductive conformal prediction is sometimes informally described as “individualized” due to its per-test-point recalibration, but the formal guarantee remains marginal.

The generalization to non-exchangeable or structured settings can alter what “transductive” means. In dynamic graphs, one regime obtains valid prediction without exchangeability by choosing the test index uniformly at random among a finite pool of node–time pairs, rather than by assuming exchangeability of the data points themselves (Davis et al., 2024). This suggests that the transductive idea is more general than the usual i.i.d. setting: validity may arise either from exchangeability of the sample or from symmetry induced by the random designation of a test point within a fixed realized pool (Davis et al., 2024).

3. Algorithmic workflow and random-forest instantiation

The conformalClassification package gives a concrete implementation of TCP for classification using random forests (Gauraha et al., 2018). In this implementation, the conformity score for observation $Y_{n+1}$9 and class $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$0 is the fraction of trees voting for class $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$1: $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$2 (Gauraha et al., 2018). This is treated as the class-specific conformity score for each point under the random-forest model (Gauraha et al., 2018).

The transductive procedure then operates label by label. For each $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$3, one forms the augmented dataset $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$4, computes conformity scores $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$5 for all points in the augmented sample, and then ranks the test-point conformity against the conformities of training examples from the same class (Gauraha et al., 2018). In the smoothed Mondrian formulation described there, the p-value is

$D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$6

where $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$7 breaks ties and $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$8 is the number of training points with label $D(y)=\{Z_1,\dots,Z_n,(X_{n+1},y)\}$9 (Gauraha et al., 2018). The paper also gives an equivalent algorithmic version counting over indices $y$0 in the augmented sample (Gauraha et al., 2018).

The key implementation detail is that, in TCP mode, the underlying model is conceptually retrained or updated for each candidate label and each test object (Gauraha et al., 2018). The paper states this directly: the fully on-line mode of TCP can be very computationally demanding and may be computationally intractable for large datasets (Gauraha et al., 2018). The 2026 overview makes the same point in more general terms, noting that full conformal requires, for each candidate $y$1, forming $y$2, fitting a model on $y$3, computing scores for all points, and evaluating the rank (Sesia et al., 25 Mar 2026).

Mondrian conditioning is built into these classwise p-values. The package paper notes that for classification problems, the error rate may be higher in some classes than others, and that applying nonconformity scores on a per-class basis is referred to as Mondrian CP (Gauraha et al., 2018). In the binary random-forest formulation for multi-source aggregation, the nonconformity scores are similarly computed only within the class corresponding to the hypothesized label, producing a smoothed Mondrian TCP (Spjuth et al., 2018). This suggests that class-conditional ranking is not merely an implementation detail but a structural response to class imbalance and class-dependent error heterogeneity.

4. Comparison with inductive conformal prediction and computational trade-offs

Transductive conformal prediction is routinely contrasted with inductive conformal prediction, also called split conformal in the recent statistical overview (Sesia et al., 25 Mar 2026). In the package paper, ICP partitions the training set into a proper training set $y$4 and calibration set $y$5, trains the underlying model once on $y$6, computes calibration scores on $y$7, and then evaluates p-values for test points by comparing their scores only to the calibration scores (Gauraha et al., 2018). The principal difference is that TCP uses the full training set plus the hypothetical test label for each candidate label, whereas ICP uses a fixed model trained once and a held-out calibration sample (Gauraha et al., 2018).

The package paper summarizes the trade-off succinctly: “TCP gives results with higher validity than ICP, however ICP is computationally faster than TCP” (Gauraha et al., 2018). There, validity is assessed empirically using deviation from validity, and TCP is described as using the full training set in a fully conformal way, whereas ICP loses some statistical efficiency because calibration is based on a smaller sample (Gauraha et al., 2018). The statistical overview frames the same comparison somewhat differently: both full/transductive and split/inductive conformal have valid marginal coverage under exchangeability, but full conformal is potentially more statistically efficient because it refits on the augmented dataset $y$8, whereas split conformal enjoys huge computational savings (Sesia et al., 25 Mar 2026). A plausible implication is that “higher validity” in the package paper refers not to a stronger asymptotic theorem than split conformal’s marginal validity, but to closer finite-sample alignment of empirical error with nominal significance under the specific diagnostics used there.

The computational burden of full conformal has motivated a growing literature on approximations. The statistical overview cites Burnaev and Vovk’s efficient conformalization for ridge regression and Lei’s piecewise-linear homotopy for lasso as examples where model structure can be exploited to reduce the cost of full conformal updates (Sesia et al., 25 Mar 2026). More recently, tournament-corrected approximations to full conformal have been proposed to restore distribution-free validity to computational shortcuts. In that framework, full conformal is explicitly identified as transductive, split conformal as inductive, and approximation schemes based on deletion, rounding, one-step updates, or approximate Bayesian posterior predictives are wrapped by a tournament correction that guarantees marginal coverage $y$9, improving to approximately $p$0 under stability conditions (Bhattacharyya et al., 28 May 2026). This line of work clarifies that the main practical obstacle to TCP is not the definition of its scores but the cost of repeatedly simulating the test point as part of the training sample.

The medical VLM literature provides a different perspective on the transductive–inductive distinction. “Transductive split conformal adaptation” uses an unsupervised transductive adaptation jointly on calibration and test inputs before applying standard split conformal calibration, precisely because direct adaptation on calibration labels breaks exchangeability between calibration and test for SCP (Silva-Rodríguez et al., 20 Jun 2025). This is not full TCP in the original Vovk–Gammerman–Shafer sense, but it inherits the central transductive idea that unlabeled test inputs may be used symmetrically with calibration inputs without sacrificing marginal validity (Silva-Rodríguez et al., 20 Jun 2025). A similar logic underlies Conf-OT for zero-shot VLMs (Silva-Rodríguez et al., 30 May 2025) and LATA for medical VLMs (Bozorgtabar et al., 19 Feb 2026), where deterministic transductive transformations of the joint calibration–test pool preserve exchangeability and hence split-conformal guarantees.

5. Efficiency, diagnostics, and known limitations

The conformalClassification package paper defines several diagnostics for assessing conformal predictors (Gauraha et al., 2018). For a test set $p$1 and prediction regions $p$2, the error rate is

$p$3

efficiency is

$p$4

deviation from validity is

$p$5

and observed fuzziness is

$p$6

(Gauraha et al., 2018). These metrics separate the issues of validity, ambiguity, and concentration of mass on incorrect labels.

The distinction between coverage and efficiency is central in the transductive literature. Full conformal is often regarded as more efficient than split conformal because it uses all available data in a label-specific recalibration, but efficiency is never automatic: the quality of the score function is decisive (Sesia et al., 25 Mar 2026). The statistical overview makes this explicit by pointing out that conformal validity holds for any score $p$7, but poor score design can yield trivial or uninformative sets (Sesia et al., 25 Mar 2026). In the conformalClassification setting, the raw random-forest vote proportion is the chosen conformity score, with no additional transformation (Gauraha et al., 2018).

A recent information-theoretic analysis studies a stricter transductive setting in which the prediction target is a whole vector of $p$8 test labels rather than one label at a time (Behboodi et al., 4 Sep 2025). There, any nontrivial joint confidence level entails exponential growth of the expected size of the prediction set in the number of test points, with exponent at least the conditional entropy $p$9, plus a second-order dispersion term (Behboodi et al., 4 Sep 2025). This is a different notion of transductive prediction from classical single-point full conformal, but it clarifies that the efficiency–confidence trade-off can become severe when one requires simultaneous correctness for multiple labels rather than marginal coverage for one test point (Behboodi et al., 4 Sep 2025).

Another limitation concerns practical computation of exact full conformal regions over continuous response spaces. In regression, determining the exact region may require evaluating the plausibility function over all $D(y)$0, so many implementations rely on discretization or numerical search (Hong et al., 11 Oct 2025). That paper argues that such approximate regions may jeopardize finite-sample validity if they differ from the exact conformal region, and proposes specially designed nonconformity measures that make the transductive conformal region analytically tractable as one-sided or two-sided intervals (Hong et al., 11 Oct 2025). This suggests that some practical challenges attributed broadly to full conformal are in fact challenges of score design and numerical representation.

6. Extensions, specialized domains, and broader interpretations

Transductive conformal prediction has been adapted to a wide range of specialized settings. In privacy-preserving multi-source prediction, each site runs its own local Mondrian TCP on undisclosed data and returns only p-values, which are then averaged label-wise to form a Non-Disclosed aggregated Conformal Predictor (Spjuth et al., 2018). The resulting procedure is not itself a standard conformal construction, so exact global validity is not guaranteed by classical conformal theory, but the paper reports conservative empirical validity for significance levels $D(y)$1 to $D(y)$2 and reduced variance relative to individual small-sample TCPs (Spjuth et al., 2018). This shows that the transductive mechanism can be deployed locally even when the final aggregation step falls outside standard conformal guarantees.

In dynamic graph learning, transductive conformal prediction takes on a different structural meaning. The “transductive regime” there assumes that the missing label is selected uniformly at random among a finite set of node–time pairs; validity then follows without exchangeability assumptions on the graph process itself (Davis et al., 2024). The paper also presents a split-conformal algorithm based on unfolded dynamic-graph representations and APS scores, and proves validity for the semi-inductive regime under exchangeability-type assumptions and algorithmic symmetry (Davis et al., 2024). This extends the transductive idea beyond i.i.d. tabular prediction into structured, dependent data.

In multivariate regression, transductive reasoning has recently been used to standardize residuals coordinate-wise without extra data splitting. There, the method is conceptually full/transductive because it reasons about how the test residual would affect the standardization parameters if it were included in the calibration pool, but the resulting algorithm is implemented with fast reusable thresholds and still guarantees finite-sample joint coverage for hyperrectangular sets (Fan et al., 17 Dec 2025). This suggests that transductive conformal ideas can be embedded within algorithms that are computationally closer to split conformal than to classical full conformal, while preserving the logic of “including the test point in the calibration geometry.”

Finally, transductive conformal prediction has been linked to imprecise probability theory. Under a consonance assumption on the conformal transducer, the conformal p-function can be interpreted as a plausibility contour, inducing a credal set whose imprecise highest density region coincides exactly with the classical conformal prediction region (Caprio et al., 10 Feb 2025). In that interpretation, transductive conformal regions become imprecise highest density regions, and consonant plausibility functions are monoid homomorphisms from $D(y)$3 to $D(y)$4 (Caprio et al., 10 Feb 2025). This provides a different conceptual lens: TCP is not only a model-free frequentist procedure with finite-sample coverage, but also a specific construction of predictive credal sets and level sets within imprecise probability theory.

Overall, transductive conformal prediction remains the canonical conformal method: exact, distribution-free, finite-sample valid under exchangeability, and conceptually elegant because it calibrates each test-label hypothesis against the entire augmented sample (Sesia et al., 25 Mar 2026). Its enduring challenges are computational cost, score design, and the gap between marginal and conditional guarantees. Much of the recent literature can be read as an effort either to preserve the statistical advantages of full/transductive conformal while reducing its computational burden (Bhattacharyya et al., 28 May 2026), or to transplant its symmetry-based logic into new settings such as dynamic graphs (Davis et al., 2024), zero-shot transfer (Silva-Rodríguez et al., 30 May 2025), and trustworthy few-shot medical adaptation (Silva-Rodríguez et al., 20 Jun 2025).