White-Box Machine Learning Algorithms

• White-box machine learning algorithms are models with transparent internal logic that enable direct inspection of decision processes using symbolic representations.
• They often combine interpretable components like shallow decision trees or polynomial networks with black-box modules to achieve high predictive performance.
• These approaches are applied in scientific computing, simulation acceleration, and human-in-the-loop systems, balancing interpretability with competitive accuracy.

White-box machine learning algorithms are defined by their transparent, interpretable internal logic, which allows direct inspection and human understanding of their decision processes. These models, in contrast to black-box methods, expose their structure or reasoning either through explicit algebraic forms, constrained architectures, or tightly physics-aware surrogates, making them prominent tools for scientific computing, engineered systems, and interpretable AI. White-box learning encompasses shallow decision trees, rule lists, algebraic polynomial networks, and structured surrogates embedding first-principles models, as exemplified in recent research addressing the trade-off between model interpretability and predictive performance (Vernon et al., 2024, Ellis et al., 8 Sep 2025, Liu et al., 2020).

1. Fundamental Principles and Defining Characteristics

The principal attribute of white-box machine learning is algorithmic transparency: the ability to trace and comprehend every step mapping inputs to outputs. This contrasts with black-box models, such as deep ensembles or large random forests, whose internal state and feature contributions are difficult or impossible to explicitly enumerate (Vernon et al., 2024). A model may be considered white-box if:

• Its structure is amenable to symbolic inspection (e.g., decision trees of bounded depth, direct polynomial mappings).
• Outputs are directly interpretable in terms of input features and weights.
• Physical or logical constraints can be embedded, verified, or enforced throughout the inference pipeline.
• Surrogate modeling only replaces sub-components rather than the entire process, preserving model integrity and domain interpretation (Ellis et al., 8 Sep 2025).

2. Canonical Architectures and Algorithms

Several architected forms exemplify white-box machine learning:

Decision Tree Ensembles with Explicit Gating

"Integrating White and Black Box Techniques for Interpretable Machine Learning" (Vernon et al., 2024) describes a three-component ensemble:

• Base classifier ($f_b$): interpretable white-box model (e.g., decision tree of depth $D=4$); handles 'easy' cases.
• Deferral classifier ($f_d$): high-capacity black-box (e.g., random forest) for 'hard' cases.
• Grader ($g$): a separate shallow decision tree routing each instance to $f_b$ or $f_d$ based on the predicted difficulty, fitted on auxiliary labels $z_i = \mathbb{I}[f_b(x_i) \neq y_i]$.

Algebraic Polynomial Networks

"Dendrite Net: A White-Box Module for Classification, Regression, and System Identification" (Liu et al., 2020) introduces the Dendrite Net (DD), an L-layer module where each layer computes

$A^{l} = (W^{l,l-1}A^{l-1}) \circ X$

(with $\circ$ the Hadamard product), resulting in an explicit multivariate polynomial expansion. This encoding yields fully interpretable symbolic mappings, distinguishing all direct and combinatorial input influences up to degree $L$.

Structured Physical Surrogates

"Fast phase prediction of charged polymer blends by white-box machine learning surrogates" (Ellis et al., 8 Sep 2025) demonstrates a surrogate structured to preserve physical interpretability. Only the computational bottleneck of the Random Phase Approximation—specifically, the form-factor matrix $\Gamma(x, k)$—is replaced by a parallel partial Gaussian process, while the remainder of the first-principles model remains intact.

Key Features Table

Model Type	Interpretability Mechanism	Typical Use Case
Shallow Decision Tree	Symbolic path inspection, bounded depth	Classification, Gating/Grading
Dendrite Net (DD)	Algebraic polynomial expansion	Regression, System ID, White-box DL module
Physics-Aware Surrogate	Blocks replaced, physical constraints	Scientific simulation acceleration

3. Mathematical Formulation and Training

White-box models prioritize forms that retain clarity in parameter–output relationships.

• Decision Trees: $h \in \mathcal{H}_{\text{DT}, D}$, with constraints $\mathrm{depth}(h) \le D$, $\#\mathrm{nodes}(h) \le 2^{D+1}-1$ (Vernon et al., 2024).
• Polynomial Networks (DD): Stacking $L$ DD modules maps $X$ to $Y$ via repeated matrix multiplications and Hadamard products, producing a high-order explicit polynomial representation. All coefficients can be symbolically extracted post-training (Liu et al., 2020).
• Surrogate Models: Surrogates for sub-models use Gaussian processes parameterized by architecture variables and informed by kernel structure consistent with physical smoothness and block structure (Ellis et al., 8 Sep 2025).

Training methodologies follow standard objective minimization:

• Classification: cross-entropy $\mathcal{L}_{\mathrm{CE}}$.
• Regression: mean-squared error $\mathcal{L}_{\mathrm{MSE}}$.
• Regularization is implicit via hard constraints (e.g., depth, module count) or explicit (e.g., $\ell_2$ penalties).

Model selection and validation are primarily executed by monitoring interpretability metrics (e.g., model depth, number of polynomial terms) alongside predictive accuracy.

4. Accuracy–Interpretability Trade-offs and Empirical Findings

Empirical studies demonstrate the central trade-off inherent to white-box learning:

• Shallow trees alone deliver lower accuracy than powerful random forests, but combining a white-box base and grader with a black-box deferral yields ensemble accuracy close to the black-box, while restricting black-box invocation to 'hard' regions only (Vernon et al., 2024).
• In system identification and regression, Dendrite Nets outperform or match multi-layer perceptrons and SVMs in test accuracy while maintaining an algebraic, inspectable form (Liu et al., 2020).
• In phase-behavior prediction, a white-box model using just $O(10^2)$ samples achieves $>$99% out-of-sample accuracy compared to $\sim$80–85% for black-box ML (requiring $O(10^3)$ or more samples) (Ellis et al., 8 Sep 2025).

Table excerpted from (Vernon et al., 2024) illustrates the ensemble impact:

Dataset	$\mathrm{Acc_{base}}$ [%]	$\mathrm{Acc_{ensemble}}$ [%]	Deferral Rate [%]
Gas	72.96	95.50	37.44
Breast	93.28	93.92	9.13
Yeast	56.85	61.16	67.90

A plausible implication is that by precisely targeting the input regions where white-box models exhibit high uncertainty or error, ensemble routing can efficiently preserve global interpretability while recovering nearly maximal accuracy.

5. Interpretability Metrics, Model Complexity, and Limitations

Interpretability in white-box machine learning is characterized by explicit, bounded model complexity, symbolic expandability, and the preservation of domain semantics:

• Decision Tree Complexity: measured by maximum depth $D$ or internal node count; hard limits are directly imposed (Vernon et al., 2024).
• Dendrite Net Expressivity: controlled by the number of modules $L$, which directly corresponds to the polynomial degree; excessive module count is naturally attenuated by small high-order coefficients, avoiding catastrophic overfitting (Liu et al., 2020).
• Physics-Aware Constraints: surrogates are trained and tested to respect physical positivity, block structure, and algebraic symmetries (Ellis et al., 8 Sep 2025).

Nevertheless, several limitations are identified:

• Gating accuracy in ensemble frameworks is not theoretically bounded; false negatives in identifying 'hard' regions lead to degraded overall performance (Vernon et al., 2024).
• Dendrite Net has not yet been scaled or validated on large-scale convolutional architectures; symbolic expansion for very high-dimensional input remains a challenge (Liu et al., 2020).
• White-box surrogates are most effective when a single computational bottleneck dominates; end-to-end white-box modeling of arbitrarily complex or high-dimensional patterns may not be achievable with tractable interpretability or accuracy (Ellis et al., 8 Sep 2025).

6. Practical Applications and Extension Across Domains

The white-box paradigm is applied in several distinct research contexts:

• Human-in-the-loop Decision Support: Systems that require direct auditability and traceability, such as medical or legal AI, leverage white-box models to ensure compliance and enable user trust (Vernon et al., 2024).
• Physics-based Simulation Acceleration: In material science, white-box surrogates for costly sub-models enable rapid and interpretable screening of design spaces without loss of first-principles fidelity (Ellis et al., 8 Sep 2025).
• General-purpose Machine Learning Modules: Dendrite Net is proposed as a building block not only for standalone regression or classification, but as a drop-in white-box module for deep, hybrid, or attention-based architectures (Liu et al., 2020).

Extension opportunities include:

• Integration with other white-box representations (rule lists, small linear models).
• User-interface advancements to further visualize and clarify white-box, black-box, and gating regions for domain experts (Vernon et al., 2024).
• Adapting white-box partial surrogates to other domains where kernelized sub-manifolds or analytic computation steps dominate computational cost or interpretability requirements (Ellis et al., 8 Sep 2025).

In summary, white-box machine learning algorithms offer a rigorously interpretable and domain-amenable alternative to purely statistical black-box models, often attaining competitive accuracy through architectural fusion, hierarchical surrogacy, or direct algebraic construction. The continued development of scalable and robust white-box architectures remains a central theme for interpretable AI, scientific machine learning, and AI safety.