Predictive Partitioning Methods

• Predictive partitioning is a framework that treats the partition as an integrated model component, optimizing it based on downstream predictive performance.
• It employs methods such as smooth approximations, gradient-based optimization, and dynamic programming to refine partition boundaries.
• It is applied in diverse fields like survival analysis, graph processing, and control systems to balance trade-offs between prediction accuracy and computational cost.

Predictive partitioning denotes a family of methods in which the partition of a space is itself selected by a predictive, inferential, or operational objective. The term is used for several related ideas: learning cut points for event-time intervals in survival analysis (Hickey et al., 2023), clustering multi-view observations so that within-cluster predictive ability is maximized (McWilliams et al., 2012), weighting graph partitions by expected Breadth First Search communication (Fay, 2015), selecting subsystem decompositions for non-centralized Model Predictive Control (MPC) to optimize control-relevant metrics (Riccardi et al., 14 Sep 2025), and choosing contiguous covariate bins to minimize leave-one-out Continuous Ranked Probability Score (CRPS) in conformal regression (Toccaceli, 23 Mar 2026). Taken together, these lines of work suggest a general principle: a partition is treated not as a fixed preprocessing choice, but as a model component whose quality is judged by its downstream predictive consequences.

1. Scope of the concept

The objects being partitioned vary substantially across the literature. In survival analysis, the partition lies on the event-time axis or on a prognostic factor; in multi-view learning, it is a partition of observations into clusters; in emulation and localized forecasting, it is a decomposition of input or time domains; in graph processing and distributed control, it is a partition of vertices, edges, or dynamical subsystems; in conformal regression, it is a set of contiguous bins over sorted covariates; and in protein representation learning, it is a partition of a residue graph into connected substructures (Hickey et al., 2023, McWilliams et al., 2012, Surjanovic et al., 2019, Fay, 2015, Riccardi et al., 14 Sep 2025, Toccaceli, 23 Mar 2026, Wang et al., 2 Oct 2025).

Domain	Partitioned object	Criterion emphasized
Survival analysis	Event-time space or prognostic factor	Likelihood, CI, AUC, log-rank separation
Multi-view learning	Observation clusters	Within-cluster predictive ability
Emulation and DMD	Input space or time domain	Local prediction error, residuals
Graphs and control	Vertices, edges, subsystems, coalitions	Message cost, stage cost, communication, computation
Conformal regression	Contiguous covariate bins	Total LOO-CRPS
Representation learning	Feature space or protein graph	Task-aware specialization or interpretable substructures

A recurring contrast is between predictive criteria and purely geometric or topological criteria. MVPP explicitly distinguishes “predictive clustering” from “geometric clustering,” arguing that points should be grouped according to how well they are explained by the same regression model rather than by geometric closeness alone (McWilliams et al., 2012). The same contrast appears in event-time prediction, where fixed quantile-based intervals are replaced by learned cut points (Hickey et al., 2023), and in control, where performance-based or predictive partitioning is contrasted with structure-only decompositions (Riccardi et al., 14 Sep 2025).

2. Statistical formulations

In survival analysis, learned event-time partitioning starts from a piecewise constant density over intervals $I_j=(c_{j-1},c_j]$, with interval probabilities predicted by a neural network:

$\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$

Because the indicator is discontinuous in the cut points, the method replaces it with a smooth approximation via sigmoids and then jointly optimizes the network parameters and the cut points under the likelihood for non-informative right censoring. The formulation includes a Beta prior on cut points, entropy regularization, and gradient-based optimization with Adam (Hickey et al., 2023).

A different statistical formulation appears in K-adaptive partitioning for survival subgrouping. For a fixed number of subgroups $K$, candidate split sets are scored by the minimum pairwise log-rank statistic,

$T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$

and the selected partition maximizes this minimum:

$T_1^*(s_K^*) = \max_{s_K \in S_K} T_1(s_K).$

The number of subgroups is then chosen by a permutation test. This criterion is explicitly designed to avoid a partition in which one pair of groups remains weakly separated even if the overall split appears strong (Eo et al., 2013).

In multi-view predictive partitioning, the central object is the predictive influence of a point under a two-block partial least squares (TB-PLS) model. The clustering objective minimizes the within-cluster sum of squared predictive influences,

$$C(\Theta,\mathcal{C}) = \sum_{k=1}^{K} \sum_{i \in \mathcal{C}_k} \left\Vert \bm{\pi}^{(k)}_{\mathbf{x},\mathbf{y}}(\mathbf{x}_i,\mathbf{y}_i) \right\Vert^2,$$

where predictive influence is defined as the total derivative of the PRESS statistic with respect to the point. This makes the partition depend on out-of-sample predictive behavior rather than on distances in either view (McWilliams et al., 2012).

CRPS-optimal binning for conformal regression adopts yet another formulation. Observations are sorted by a one-dimensional covariate, partitioned into contiguous bins, and assigned the within-bin empirical CDF as the predictive distribution. For a bin $S$ of size $m$, the total leave-one-out CRPS reduces to

$\mathrm{cost}(S) = \frac{mW}{(m-1)^2},$

where $W$ is the total pairwise absolute difference within the bin. Because the total objective is additive over bins, the globally optimal $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$0-partition can be recovered by dynamic programming in $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$1 time after $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$2 precomputation (Toccaceli, 23 Mar 2026).

These formulations differ in loss function, stochastic assumptions, and optimization method, but they share a structural idea: partition boundaries are estimated from the same predictive criterion that ultimately evaluates the model.

3. Adaptive, competitive, and differentiable partition learning

Several methods learn partitions by iteratively focusing model capacity where predictive difficulty is greatest. The adaptive partitioning emulator (APE) begins with the full input space as one region, fits a Gaussian process, identifies the region with the highest leave-one-out cross-validation error, densifies that region, and then splits it along the dimension that minimizes a within-versus-between variability criterion. Each resulting region receives an independent local Gaussian process. On the 10-dimensional corner peak function, APE with $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$3 reached scaled RMSPE $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$4 at $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$5, matching the global GP’s $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$6 at $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$7, while achieving scaled MAPE $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$8 versus $\hat{p}(t \mid \mathbf{x}) = \sum_{j=1}^{M+1} p_\phi(z_j \mid \mathbf{x}) \frac{\mathbbm{1}_{I_j}(t)}{|I_j|}.$9 for the global GP; the same study states that APE and global GP reach similar accuracy at similar $K$0, but APE’s time grows slowly while global GP becomes intractable at $K$1 (Surjanovic et al., 2019).

Active partitioning inverts the active learning paradigm by rewarding models for the points they already predict best. At each epoch, multiple models compete on each data point, the winner is trained on that point, and adding or dropping models is allowed when poorly predicted or redundant regions emerge. The resulting specializations define the partition. In the reported experiments, modular models built on these partitions achieved a 54% loss reduction on porous structure stress-strain data, 53% on the energy efficiency dataset, 14% on the automobile dataset, and 8% on students’ grades; the same summary states that on well-mixed or unimodal datasets the modular model offers less or no advantage (Tacke et al., 2024).

Localized Dynamic Mode Decomposition applies a related idea in the time domain. The time interval is segmented into stages, DMD is fit independently within each stage, and an adaptive partitioning strategy starts a new stage when a residual-based error estimator exceeds a tolerance. The paper reports orders-of-magnitude improvements in mean relative error: for Burgers’ equation, DMD yielded $K$2 while LDMD achieved $K$3 to $K$4; for the Allen–Cahn equation, DMD gave $K$5 and LDMD $K$6 (Li et al., 17 Mar 2025).

Differentiable partitioning appears in representation learning as well. BioBlobs partitions a protein graph into flexibly-sized, non-overlapping, connected substructures through a seed-selection and blob-expansion mechanism using Gumbel-Softmax, straight-through estimators, threshold prediction, and vector quantization into a shared codebook. The abstract states that these representations improve the performance of widely used protein encoders such as GVP-GNN across various protein representation learning tasks, while also yielding an interpretable vocabulary of function-relevant substructures (Wang et al., 2 Oct 2025).

This range of mechanisms shows that predictive partitioning is not tied to a single optimization paradigm. Gradient relaxation, dynamic programming, local error refinement, competitive specialization, and residual-triggered repartitioning all instantiate the same design choice from different computational angles.

4. Interpretability, explanation, and personalized inference

Predictive partitioning is frequently used not only to improve accuracy but also to produce structured explanations. RIPE extracts a sparse set of hyperrectangles, interpretable as If-Then rules, and then turns this overlapping rule set into a partition of feature space. Prediction is made as the empirical mean of the target among training observations lying in the same partition cell. The method is deterministic, uses a sparse rule set, and in a $K$7 feature example reportedly used only 14 rules, mostly involving 2–3 features (Margot et al., 2018).

TreeView approaches interpretability from the side of learned neural representations. Hidden neurons are clustered into factors, samples are clustered within each factor, the resulting cluster labels are assembled into meta-features, and a decision tree is trained on those meta-features. The summary reports a 96% accuracy for the original DNN and 94% for the TreeView surrogate on a UCI Image Segmentation case study, while retaining a hierarchical view of class-hypothesis rejection across factors (Thiagarajan et al., 2016).

Cross-validated variable selection in tree-based methods places prediction directly inside the splitting rule. ALOOF selects the splitting variable at each node by leave-one-out cross-validation loss and then performs a standard CART split only for the selected variable. This removes the bias toward categorical variables with many categories. The summary states that for two-class classification the method can be designed so as not to increase the overall computational complexity compared to CART under reasonable assumptions, whereas for regression numerical splitting becomes $K$8 rather than $K$9 (Painsky et al., 2015).

In personalized treatment selection, product partition models with covariates use predictive biomarkers to encourage clustering, while prognostic covariates enter the response model. The resulting NGGP-PPMx framework performs clustering and treatment-response prediction jointly, propagating partition uncertainty into treatment selection. The summary reports that the method consistently outperformed competing heuristic clustering-plus-prediction procedures on Number of Optimal Treatment misassignments, gain in treatment utility, and number of patients with correctly predicted outcomes, especially under substantial heterogeneity (Pedone et al., 2022).

A plausible implication is that predictive partitioning often serves a dual role: it organizes statistical dependence in a way that improves out-of-sample behavior, and it externalizes structure that can be inspected as rules, clusters, factors, or clinically meaningful subgroups.

5. Graphs, infrastructures, and non-centralized predictive control

In graph processing, predictive partitioning has been used to reduce communication during parallel Breadth First Search. The method models the expected degree distribution in the BFS frontier, derives expected edge-use probabilities for a target iteration, assigns these as edge weights, and uses METIS to minimize weighted cross-partition traffic. The abstract and results summary report reductions of up to 50% in some instances and in the order of 20% on average in social networks, while also noting no improvement for Erdős–Rényi graphs (Fay, 2015).

In non-centralized control, predictive partitioning becomes a system-design problem. Event-triggered repartitioning for interconnected microgrids divides the network into non-overlapping connected microgrids, triggers repartitioning when at least one microgrid becomes non-self-sufficient, and then forms coalitions among neighboring microgrids when self-sufficiency cannot be achieved. The method guarantees feasibility, produces a feasible but sub-optimal solution to the centralized problem, reported average suboptimality over a day is 21.93%, and communication is limited to within coalitions, with no communication required when all microgrids are self-sufficient (Ananduta et al., 2020).

Dynamic clustering for distributed MPC in solar parabolic trough plants provides a more performance-oriented variant. Loops are periodically re-clustered from features $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$0, the aggregated optimization is solved with ALADIN, and the number of coordinated variables is reduced when loop behavior is similar. In the 10-loop case study, the finest static partition achieved cumulative cost $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$1, the coarsest $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$2, and a dynamic partition with $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$3 achieved $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$4–$T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$5, with substantially fewer coordinated variables (Chanfreut et al., 2023).

A more general theory is developed in the framework of Fundamental System Units (FSUs), Composite System Units (CSUs), and a partition index that balances intra-CSU interaction, inter-CSU interaction, and CSU size through a granularity parameter $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$6. The case studies involve modular and random networks with 64 and 50 FSUs, and the reported outcome is significant reductions in computation time and cost while maintaining or improving control performance with respect to conventional strategies (Riccardi et al., 28 Feb 2025). The 2025 survey systematizes this area into five classes—optimization-based, algorithmic, community-detection-based, game-theoretic-oriented, and heuristic approaches—and formalizes predictive partitioning as dynamic determination of subsystem definitions over the MPC horizon using graph-theoretical representations and quality metrics such as cumulative stage cost, computation time, computation cost, and communication cost (Riccardi et al., 14 Sep 2025).

Predictive partitioning also appears in database design as workload-aware horizontal fragmentation. There, the objective is to choose fragments that minimize the total execution cost of a query workload, predicted by altering PostgreSQL catalog statistics rather than loading real data. A genetic algorithm searches the space of candidate schemas, and the initial evaluation reports a reduction of estimated query cost from approximately 2 million to less than 200,000 disk accesses, mean estimation error of 1.16% with maximum error 3.10%, and average real execution time reduced from 11,168 ms to approximately 1270 ms (Arsov et al., 2019).

6. Evaluation criteria, failure modes, and recurrent controversies

A central empirical result in learned event-time partitioning is that, across the German Breast Cancer Study Group, FL Chain, and Pooled Stroke Risk Cohorts datasets, learning cut points provided higher or equal CI and AUC than the baseline across all datasets and numbers of cut points, with the largest improvement in the smallest data sets; the same summary states that best performance was usually with only 3 learned intervals and that increasing the number of intervals did not help the baseline but did not degrade the learned model as much (Hickey et al., 2023). This directly challenges the common practice of fixing time intervals a priori.

The multi-view literature raises a related warning against structure-only clustering. MVPP is specifically designed for cases in which geometric structure is misleading but predictive relationships between views remain cluster-specific; the reported simulations distinguish a geometric scenario, where many methods work, from a predictive or confounding scenario, where MVPP dramatically outperforms geometric alternatives (McWilliams et al., 2012). In graph partitioning, the analogous caveat is that predictive weighting helps when graph structure is strongly skewed, but not for structureless Erdős–Rényi graphs (Fay, 2015).

Model-selection criteria themselves can be problematic. CRPS-optimal binning shows that minimizing within-sample LOO-CRPS is inappropriate for selecting the number of bins $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$7 because it results in in-sample optimism; the method therefore selects $T_1(s_K) = \min_{1 \le g < h \le K} \chi_1^2(g,h),$8 by test CRPS on an alternating held-out split, which yields a U-shaped criterion with a well-defined minimum. On real benchmarks against Gaussian split conformal, CQR, and CQR-QRF, the method is reported to produce substantially narrower prediction intervals while maintaining near-nominal coverage (Toccaceli, 23 Mar 2026).

Control-oriented partitioning introduces further complications because quality depends on the metric used to judge it. In district heating networks, a nominally optimal partition exhibited an average cost increase of only 2.8% relative to centralized control across eleven of the twelve perturbed cases, and in three cases the nominal partition remained globally optimal. Yet the same study reports that in five of twelve cases the case-specific OLM-minimizing partitions underperformed the nominally optimal one because of shifts in the relative magnitude of heat loss versus flexibility costs; proper tuning of cost-function weights and initial conditions was therefore identified as essential, and seasonal repartitioning was recommended when demand profiles deviated substantially from nominal, as in the November operating case (Blizard et al., 11 May 2026).

The control survey generalizes this point by stressing that no single partition necessarily minimizes stage cost, computation time, computation cost, and communication cost simultaneously (Riccardi et al., 14 Sep 2025). A plausible implication is that predictive partitioning is best understood not as a universal recipe for finer discretization, but as a framework for exposing and optimizing trade-offs that would otherwise remain hidden behind fixed partitions.