Component-Based Machine Learning

• Component-Based Machine Learning (CBML) is a design paradigm that structures models as explicit, modular components to enhance interpretability and maintenance.
• It decomposes tasks using diverse modules—such as pipeline stages, concept predictors, and residual learners—to improve reusability, adaptability, and data efficiency.
• CBML facilitates system-level assurance by providing clear interfaces for editing, monitoring, and replacing model parts, which supports targeted interventions and robust performance.

Component-Based Machine Learning (CBML) denotes a family of modeling practices in which a machine-learning system is organized as a composition of modular components with explicit semantics, interfaces, and composition rules rather than as a single monolithic predictor. Across the literature, components may be pipeline stages, additive base learners, concept predictors, interaction modules, latent aggregators, knowledge-based simulators, explainers, safety monitors, or editable submodels. The shared premise is that decomposition can improve reuse, controllability, interpretability, maintenance, and in some settings data efficiency or robustness, while preserving competitive task performance (Coors et al., 2021, Du et al., 10 Nov 2025, Chen et al., 2023, Wang et al., 25 Jul 2025, McCarthy et al., 2024).

1. Conceptual basis

A recurring definition of CBML is the construction of learning systems from “modular, well-defined components” that are “composable and interpretable” (Coors et al., 2021). In the concept-bottleneck literature, the same idea is expressed as building systems from “interchangeable parts with well-defined interfaces,” so that maintenance, upgrade, and reconfiguration can proceed “without end-to-end retraining” (Du et al., 10 Nov 2025). In building engineering, CBML is described as a “knowledge-encoded, data-driven modeling approach” in which “the abstraction of building structural knowledge is encoded as semantic information in the model’s organization” (Chen et al., 2023).

This component view is broader than a choice of architecture. In some works, the decomposition is epistemic: the system is partitioned according to known scientific structure, such as recurring building elements, subgrid-scale closure versus resolved dynamics, or frequency-domain component structure in microwave-sky separation (Chen et al., 2023, McCarthy et al., 2024). In others, it is representational: objects, concepts, residual concepts, latent flow features, or additive partial effects become the units of composition (Shang et al., 2024, Steinmann et al., 30 May 2025, Wang et al., 25 Jul 2025). A plausible implication is that CBML is best understood not as a single model class but as a design pattern for aligning model structure with domain structure.

Several papers explicitly distinguish CBML from adjacent ideas. In building energy prediction, CBML “goes beyond crafting features” because it restructures “the model itself around domain entities and their physical relationships,” and it differs from physics-informed ML because physics is encoded “through decomposition and composition” rather than by adding constraints inside one network (Chen et al., 2023). In signal-preserving CMB separation, the hybrid estimator retains a constrained linear baseline and restricts ML to signal-free combinations, so the learned component is not allowed to absorb the target signal (McCarthy et al., 2024). These distinctions matter because they locate CBML at the level of system organization, not merely at the level of feature engineering or post-hoc explanation.

2. Component types and interface design

CBML systems differ primarily in what they treat as components. In interpretable AutoML, the decomposition appears simultaneously at the pipeline and model levels. Pipeline components include preprocessing, removal of constants, factor-level collapsing, missing-value imputation, Hyperband-based model selection and tuning, and evaluation. Model components are additive base learners: linear and centered nonlinear univariate effects, plus restricted pairwise interaction learners selected by componentwise boosting (Coors et al., 2021).

Concept-bottleneck systems expose a different interface. The canonical split is $g: X \rightarrow C$ for concept prediction and $f: C \rightarrow Y$ for label prediction, yielding $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$ (Du et al., 10 Nov 2025). Flexible Concept Bottleneck Model (FCBM) refines this into a concept generator or annotator, a text encoder, an image encoder, a concept predictor $g$, a hypernetwork $H$ that maps concept embeddings to class weights, a sparse selector based on temperature-controlled sparsemax, and a linear decision layer (Du et al., 10 Nov 2025). Incremental Residual Concept Bottleneck Model (Res-CBM) augments a base concept bank with residual vectors $U$, a residual classifier $\Psi_r$, and an incremental concept discovery module that promotes residual vectors into named concepts aligned with a candidate concept bank (Shang et al., 2024). Object-Centric Concept Bottlenecks (OCB) shift the unit of modularity from the whole image to object proposals, concept extraction on crops, aggregation functions such as max, sum, count, or concat, and a linear predictor (Steinmann et al., 30 May 2025). Controllable Concept Bottleneck Models (CCBMs) make the concept extractor $g$, concept bottleneck, and linear label predictor $f$ explicitly editable components with APIs for gradients, curvature approximations, and concept-set manipulation (Lin et al., 1 Jan 2026).

Scientific and engineering CBML instantiations are equally modular but often around physically meaningful subsystems. For pure-component property estimation, the pipeline is split into a molecular representation module based on the connectivity matrix, a supervised feature ranking and pooling module using random forest and adjusted $R^2$, interchangeable ANN or GPR predictors, and a SHAP explainer (Jiao et al., 14 May 2025). For indoor flow and temperature fields, the surrogate is composed of a convolutional autoencoder with residual connections (CAER), an MLP mapping inlet conditions to latent representations, and a CNN aggregator that combines single-inlet latents into dual-inlet fields (Wang et al., 25 Jul 2025). In QML, the components are feature maps $f: C \rightarrow Y$0, variational circuits $f: C \rightarrow Y$1, measurements $f: C \rightarrow Y$2, optimizers, kernels, and the quantum–classical loop $f: C \rightarrow Y$3, each with a narrow interface and its own explainability instrumentation (White et al., 14 Jun 2025).

CBML can also extend beyond predictive structure into assurance structure. In the SMIRK pedestrian AEB demonstrator, the ML-based pedestrian recognition capability is treated as a component inside a larger system containing sensors, radar logic, anomaly detection, an uncertainty manager, a rule engine, and a brake manager. The component is specified by contracts, interfaces, timing, thresholds, lifecycle artifacts, and stage-wise assurance arguments under AMLAS and SOTIF (Borg et al., 2022). This use of “component” is operational rather than representational, but it is consistent with the same modular logic.

3. Formal composition patterns

Although CBML spans distinct application areas, several recurrent formal patterns appear.

One is additive composition. In autocompboost, the prediction rule is explicitly decomposed as

$f: C \rightarrow Y$4

with univariate effects, pairwise interactions, and an optional deep-tree remainder fitted stagewise by componentwise gradient boosting (Coors et al., 2021). Here, the components are base learners selected on pseudo-residuals under equal degrees of freedom, which makes the additive structure directly inspectable.

A second is bottleneck factorization. In concept bottleneck models the core form is

$f: C \rightarrow Y$5

while FCBM replaces a fixed $f: C \rightarrow Y$6 by generated weights $f: C \rightarrow Y$7 derived from concept embeddings $f: C \rightarrow Y$8, followed by sparse linear aggregation over concepts (Du et al., 10 Nov 2025). Res-CBM further modifies the bottleneck with residual concepts,

$f: C \rightarrow Y$9

so that missing concept coverage is represented as an additive residual concept channel rather than hidden inside the backbone (Shang et al., 2024).

A third is hybrid residual correction. In signal-preserving CMB component separation, a constrained ILC estimate $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$0 is combined with an ML-predicted residual $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$1 trained only on signal-free combinations, producing

$x \rightarrow c = g(x) \rightarrow \hat y = f(c)$2

The key CBML move is not the CNN alone, but the separation between a signal-preserving linear estimator and a learned residual module confined to the null-space of the target spectral response (McCarthy et al., 2024).

A fourth is latent aggregation. In indoor airflow prediction, each component $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$3 with boundary input $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$4 is encoded as

$x \rightarrow c = g(x) \rightarrow \hat y = f(c)$5

so that single-inlet latent representations are reusable components and the aggregator learns their nonlinear interaction in dual-inlet settings (Wang et al., 25 Jul 2025). OCB uses an analogous pattern over object-level concept vectors, differing only in the nature of the components and the aggregation operator (Steinmann et al., 30 May 2025).

A fifth is closed-form component editing. CCBM uses influence-function updates and EK-FAC curvature approximations to edit concept labels, remove concepts, forget data, or add data without retraining from scratch. The paper’s generic influence-function foundation is

$x \rightarrow c = g(x) \rightarrow \hat y = f(c)$6

which is specialized separately for the concept predictor and the linear label predictor under each editing operation (Lin et al., 1 Jan 2026). This is CBML at the maintenance level: the model is decomposed into editable semantic units, not merely trainable layers.

4. Interpretability, controllability, and assurance

A central claim in CBML research is that interpretability can be built into the components rather than reconstructed afterward. In autocompboost, the partial effects are the additive component functions themselves: $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$7 for univariate terms and $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$8 for pairwise interactions. Under the additive approximation, the partial dependence function coincides with $x \rightarrow c = g(x) \rightarrow \hat y = f(c)$9 up to centering. The same framework defines risk-based feature importance $g$0 and stagewise complexity fractions $g$1, allowing interpretation of both variable relevance and how much explanatory burden falls on interpretable versus non-interpretable stages (Coors et al., 2021).

In molecular property estimation, interpretability is preserved because the representation module outputs labeled submatrix-count features rather than opaque embeddings, and SHAP is applied after feature pooling to expose which structural features contribute to $g$2, $g$3, or $g$4. The paper reports, for example, that $g$5, $g$6, and $g$7 are influenced by different structural features and that the SHAP patterns align with mechanistic interpretations such as chain length, aromaticity, and van der Waals effects (Jiao et al., 14 May 2025).

Concept-bottleneck CBML offers a more intervention-ready form of interpretability. FCBM keeps the decision layer linear in sparse concept activations and supports complete concept set replacement through a hypernetwork and distribution alignment, so the concept layer remains the explicit interface between semantics and prediction (Du et al., 10 Nov 2025). Res-CBM goes further by measuring descriptive efficiency through

$g$8

thereby making concept count and concept length part of the interpretability criterion rather than treating accuracy alone as sufficient (Shang et al., 2024). OCB exposes explanations at object granularity, although its fixed-size aggregators trade away exact object provenance for efficiency (Steinmann et al., 30 May 2025).

CCBM reframes interpretability as editability. It supports concept-label-level, concept-level, and data-level interventions through closed-form approximations, so practitioners can remove erroneous concepts, correct annotations, or unlearn data while keeping the semantic bottleneck explicit (Lin et al., 1 Jan 2026). This suggests a stricter notion of transparency: a representation is not merely inspectable if it is also operationally controllable.

Safety-oriented CBML extends interpretability into assurance. SMIRK assigns the ML component explicit system requirements such as SYS-ML-REQ1 and SYS-ML-REQ2, quantitative performance targets such as accuracy, false negative rate, FPPI, position error, and latency, and structured evidence across scoping, requirements, data, learning, verification, and deployment. The “safety cage” formed by OOD detection, the uncertainty manager, and rule-engine heuristics is itself a component composition designed to monitor and constrain the behavior of the learned detector (Borg et al., 2022).

5. Representative implementations and empirical behavior

The published instantiations of CBML span interpretable AutoML, concept-bottleneck vision, scientific surrogate modeling, molecular property prediction, signal separation, and safety-critical perception.

Domain	Componentization	Reported empirical pattern
Interpretable AutoML	Pipeline stages plus additive base learners	Competitive on many tabular datasets; weaker on image-like or strongly multiclass tasks
Concept-bottleneck vision	Concept modules, residual concepts, object proposals, hypernetwork heads	Accuracy comparable to or above prior CBMs with controlled sparsity and editable concepts
Building energy	Component models for roof, floor, wall-window, infiltration, zones	More robust than monolithic ML under extreme sparsity
Indoor flow surrogate	CAER, MLP latent mapper, CNN aggregator	High $g$9 and near-real-time inference
Molecular property estimation	Connectivity encoder, RF pooling, ANN/GPR, SHAP	Strong RMSE reductions versus GC baselines
CMB separation	ILC baseline plus signal-free ML residual	Lower residual variance than ILC while preserving target signal

For interpretable AutoML, autocompboost was evaluated on 29 small-to-medium OpenML benchmark datasets against auto-sklearn, H2O-AutoML, TPOT, Auto-WEKA, tuned and untuned random forests, and tuned glmnet under a 1-hour training budget per outer CV fold. The reported AUCs include adult at 0.911 for ACWB_no_HPO versus 0.930 for auto-sklearn and 0.926 for H2O, nomao at 0.989 versus 0.996 for auto-sklearn/H2O, and kr-vs-kp at 0.999 for ACWB_deep near the best 1.000. On multiclass log-loss, interpretable models lag, as in Fashion-MNIST with ACWB_no_HPO 0.615 versus H2O 0.294 and auto-sklearn 0.354 (Coors et al., 2021).

In concept-bottleneck vision, FCBM is evaluated on CIFAR10, CIFAR100, CUB, Places365, and ImageNet with accuracy and the average number of effective concepts $H$0. With ResNet50 and ViT-L/14 backbones and $H$1, the method is reported to match or exceed LF-CBM and CF-CBM on most datasets while maintaining similar sparsity, and after a complete concept-pool swap it generalizes “with just a single epoch of fine-tuning” (Du et al., 10 Nov 2025). Res-CBM reports, for example, CIFAR-10 accuracy $H$2 with $H$3 and CUE $H$4, compared with Lf-CBM accuracy $H$5 and CUE $H$6, and CIFAR-100 accuracy $H$7 with $H$8 and CUE $H$9 (Shang et al., 2024). OCB reports average performance over five seeds of $U$0 mAP on PASCAL-VOC, $U$1 on COCO(h), $U$2 on COCO(l), $U$3 accuracy on SUN397, and $U$4 on COCOLogic for OCB(RCNN), exceeding both CBM and equal-capacity CBM baselines on all five datasets (Steinmann et al., 30 May 2025).

For sparse-data engineering, the building-energy study reports that at 0.0125% sampling the monolithic model falls to $U$5 on Box/Random/Representative test sets, while CBML remains at $U$6. The Wall & Window component becomes highly fragile under extreme sparsity, including $U$7 in Representative, yet the building-level CBML output remains substantially more robust than the monolithic baseline (Chen et al., 2023).

For indoor airflow and temperature prediction, the CBML surrogate achieves testing velocity-magnitude $U$8 values of $U$9 and temperature $\Psi_r$0 values of $\Psi_r$1. The paper states that for 95% of the flow field the maximum absolute error is below $\Psi_r$2 m/s for velocity magnitude and $\Psi_r$3 for temperature, with inference latency on the order of 0.1 seconds per sample (Wang et al., 25 Jul 2025).

For pure-component physicochemical properties, the framework reduces the raw feature space from 13,316 to about 100 features without compromising accuracy and reports test-set RMSE reductions versus GC baselines of up to 83.8%. Reported GPR-CM test results include $\Psi_r$4 RMSE $\Psi_r$5 K and $\Psi_r$6, $\Psi_r$7 RMSE $\Psi_r$8 cc·mol$\Psi_r$9 and $g$0, $g$1 RMSE $g$2 K and $g$3, and $g$4 RMSE $g$5 bar and $g$6 (Jiao et al., 14 May 2025).

For CMB component separation, the hybrid method reports lower-variance maps than ILC, including “reducing the variance of the B-mode residual by factors of up to 5,” while retaining unbiasedness in cross-power with the target signal (McCarthy et al., 2024). For safety-critical perception, SMIRK reports that in 25 pedestrian scenarios AEB triggered and braked at the first possible frame, 23 of 25 collisions were avoided, none of the 13 OOD scenarios caused ghost braking, and the deployed detector achieved median inference time about 22.0 ms with a maximum observed 51.6 ms (Borg et al., 2022).

6. Limitations, misconceptions, and open directions

CBML does not guarantee full interpretability. Autocompboost explicitly includes an optional stage-3 deep-tree remainder to capture non-smooth, higher-order patterns, but this “decreases interpretability,” and high $g$7 means stage-1 and stage-2 effects should be interpreted cautiously (Coors et al., 2021). FCBM likewise yields sparse concept-to-class weights, but its paper cautions that CLIP similarities “are not causal,” that concept quality depends on meaningful visually grounded phrases, and that large semantic shifts across VLM versions may require more fine-tuning or re-prompting despite distribution alignment (Du et al., 10 Nov 2025).

Concept completeness and concept quality remain unresolved bottlenecks. Res-CBM is motivated by the difficulty of constructing a comprehensive concept bank, and its incremental discovery can suffer if the candidate bank is too small, causing local optima, or too large, causing oscillations; fine-grained domains remain challenging (Shang et al., 2024). OCB improves object-level reasoning, but its aggregation functions impose a trade-off: max, sum, count, and sum+count are fixed-size and efficient yet “lose which-object provenance,” whereas concat preserves provenance but scales linearly with $g$8 and can hurt performance (Steinmann et al., 30 May 2025).

In scientific and engineering settings, CBML inherits the limitations of its chosen decomposition. The building-energy study states that monolithic models may outperform CBML when data are dense and comprehensive within a single fixed structure, and that CBML depends on four prerequisites, including structured information representable by domain knowledge and a model capable of exploring the task’s high-dimensional hyperspace (Chen et al., 2023). The indoor-flow surrogate assumes a fixed 2D geometry, steady-state incompressible RANS, fixed boundary types, and no physics-informed constraints during training, so domain shift to new rooms or more complex physics is not addressed (Wang et al., 25 Jul 2025).

Maintenance-oriented CBML is also approximate rather than exact. CCBM relies on influence-function approximations, EK-FAC curvature surrogates, damping, and in some cases a linear label predictor. Its reported utility tracks retraining closely, but the paper notes that approximation error grows with edit magnitude and that “very large or distribution-shifting edits may still necessitate periodic full retraining” (Lin et al., 1 Jan 2026). Quantum CBML faces a related issue: fidelity-based pseudo-models, spectral kernel diagnostics, and component-wise SHAP or ALE provide local explanations for feature maps, ansätze, kernels, and decision surrogates, but the paper explicitly notes residual black-box elements, sampling overhead, and variance from shot noise and stochasticity (White et al., 14 Jun 2025).

A common misconception is that CBML is simply modular software. The literature instead treats modularity as epistemically meaningful: the modules are intended to correspond to additive effects, concepts, objects, physical subsystems, spectral components, or safety contracts. Another misconception is that decomposition always improves performance. The empirical record is more conditional: CBML is often strongest when domain structure is real and the interfaces are well chosen, but it can underperform black-box alternatives on highly non-smooth tasks, image-like covariates, strongly multiclass settings, or regimes where the chosen componentization omits important interactions (Coors et al., 2021, Chen et al., 2023, Du et al., 10 Nov 2025).

Across the surveyed work, the most durable research direction is not a single architecture but a systems principle: components should be semantically typed, quantitatively diagnosable, and operationally replaceable. The papers operationalize that principle through additive boosting libraries, editable concept bottlenecks, object-centric aggregation, hybrid scientific surrogates, signal-preserving residual learners, and safety-cased ML subsystems, collectively defining CBML as a program for structuring learning systems around explicit components rather than hiding structure inside a monolith.