Abstract
One approach for interpreting black-box machine learning models is to find a global approximation of the model using simple interpretable functions, which is called a metamodel (a model of the model). Approximating the black-box with a metamodel can be used to 1) estimate instance-wise feature importance; 2) understand the functional form of the model; 3) analyze feature interactions. In this work, we propose a new method for finding interpretable metamodels. Our approach utilizes Kolmogorov superposition theorem, which expresses multivariate functions as a composition of univariate functions (our primitive parameterized functions). This composition can be represented in the form of a tree. Inspired by symbolic regression, we use a modified form of genetic programming to search over different tree configurations. Gradient descent (GD) is used to optimize the parameters of a given configuration. Our method is a novel memetic algorithm that uses GD not only for training numerical constants but also for the training of building blocks. Using several experiments, we show that our method outperforms recent metamodeling approaches suggested for interpreting black-boxes.
Introduction
As machine learning models grow in complexity, they increasingly function as "black boxes," where the logic behind a prediction is obscured. In high-stakes domains like healthcare or finance, practitioners require more than just accuracy; they need to understand the underlying decision-making process.
One effective strategy is the creation of a "metamodel"—a globally interpretable surrogate model that approximates the black box. While previous efforts have utilized Symbolic Regression or Meijer G-functions, these often suffer from scaling issues or result in mathematically dense expressions that are difficult for humans to parse. This paper introduces SMPF (Symbolic Metamodeling using Primitive Functions), a method designed to generate accurate, human-readable mathematical equations with high computational efficiency.
Methodology
SMPF is built upon two primary pillars that bridge the gap between discrete symbolic search and continuous optimization:
- Kolmogorov Superposition Theorem (KST) Approximation: KST posits that multivariate functions can be represented as a composition of univariate functions. SMPF implements this using a structured, three-layer tree architecture. Unlike methods using complex Meijer G-functions, the edges of this tree represent simple, parameterized "primitive" functions such as polynomials, exponentials, and trigonometric functions.
- Memetic Algorithm (GP + GD): The search for the optimal metamodel uses a hybrid approach. Genetic Programming (GP) explores the discrete space of tree architectures through crossover and mutation. Simultaneously, Gradient Descent (GD) optimizes the continuous parameters of the primitive functions. A key innovation here is using GD to train the building blocks themselves, rather than just the leaf constants.
Experimental Setup
The authors evaluated SMPF across several benchmarks:
- Synthetic Function Approximation: Four 2-variable mathematical functions designed to test the ability to recover specific functional forms.
- Feature Selection Tasks: XOR, Nonlinear Additive, and Feature Switching datasets (10 features each) used to evaluate instance-wise feature importance.
- Real-world Data: The UCI Yacht dataset, where SMPF approximated Multilayer Perceptron (MLP) and Support Vector Machine (SVM) black boxes.
Key Results
SMPF demonstrated significant advantages over existing symbolic methods (SM, SP, and SR) and local explainers (SHAP, LIME):
- Accuracy and Fit: SMPF achieved the lowest Mean Squared Error (MSE) and highest $R^2$ scores on the majority of synthetic tasks. Notably, it successfully recovered exact functional forms, such as identifying exponential relationships that other methods missed.
- Feature Importance: In identifying relevant features for XOR and Nonlinear Additive datasets, SMPF competed favorably with popular local explainers while providing a global perspective.
- Efficiency: SMPF achieved a 4.7x speedup over Symbolic Pursuit (SP), completing training in 45 minutes versus 215 minutes. This efficiency stems from the lower computational cost of evaluating primitive functions compared to Meijer G-functions.
- Regularization: On the Yacht dataset, the symbolic metamodels produced by SMPF actually outperformed the original black-box models on test data, suggesting that the simpler symbolic forms provided a beneficial regularizing effect against overfitting.
Discussion and Future Work
By restricting building blocks to simple primitives, SMPF provides closed-form equations that domain experts can directly interpret. The use of KST as a structural prior allows the model to capture complex interactions without an explosive search space.
However, as a heuristic search, GP does not guarantee a global optimum, and the method remains sensitive to hyperparameters defining the tree structure. Future work may involve exploring deeper tree architectures, a wider library of primitive functions, and advanced non-convex optimization techniques to further refine the parameter tuning phase.
Summary updated