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 estimate instance-wise feature importance, understand the functional form of the model, and analyze feature interactions. In this work, we propose a new method for finding interpretable metamodels. Our approach utilizes the 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—particularly deep neural networks—grow in complexity, they increasingly function as "black boxes." While these models achieve high predictive accuracy, their internal logic remains opaque, posing significant challenges in high-stakes fields like healthcare or finance where understanding the why behind a decision is as critical as the decision itself.
One promising solution is metamodeling: the process of creating a simpler, "glass-box" surrogate model that mimics the behavior of the complex black-box. However, previous attempts at symbolic metamodeling often relied on complex mathematical constructs like Meijer G-functions. While these are flexible, they are often as difficult for humans to interpret as the original neural networks. This paper introduces Symbolic Metamodeling using Primitive Functions (SMPF), a method designed to find the optimal balance between mathematical flexibility and human readability.
Methodology
The Kolmogorov Blueprint
SMPF is grounded in the Kolmogorov Superposition Theorem (KST). KST provides a theoretical guarantee that any continuous multivariate function can be represented as a finite composition of univariate functions and additions. SMPF leverages this by representing the metamodel as a three-layer tree structure:
- Input Layer: Receives the raw features.
- Middle Layer: Computes sums of "inner" univariate functions.
- Output Layer: Applies an "outer" univariate function to produce the final prediction.
A Memetic Search Algorithm
To find the most effective and interpretable tree structure, the authors propose a hybrid "memetic" algorithm that combines global and local search strategies:
- Genetic Programming (GP): This evolutionary component searches the discrete space of tree architectures. It determines which features connect to which nodes and selects the most appropriate "primitive functions" (e.g., $\sin(x)$, $e^x$, $\log(x)$, or polynomials) for each edge.
- Gradient Descent (GD): Once the GP proposes a structure, GD is employed to fine-tune the continuous numerical parameters (coefficients and offsets) within the chosen primitive functions.
This dual-optimization approach allows the model to simultaneously discover the underlying logic (the functional form) and the precise fit (the parameters).
Experimental Evaluation
The researchers evaluated SMPF against several benchmarks, including Symbolic Pursuit (SP) and standard Symbolic Regression (SR), across three distinct categories:
- Synthetic Recovery: Testing if the algorithm could rediscover known mathematical ground truths (e.g., $e^{-3x_0+x_1}$).
- Feature Selection: Using datasets like XOR and Nonlinear Additive to determine if the model could correctly identify relevant features amidst noise.
- Real-World Surrogates: Explaining black-box models (MLPs and SVMs) trained on UCI datasets, such as Yacht Hydrodynamics and Energy Efficiency.
Key Insights and Results
Performance and Accuracy
SMPF demonstrated superior performance in approximating complex functions. In synthetic tests, it consistently achieved $R^2$ scores near 0.99. When approximating an MLP trained on the Yacht dataset, SMPF achieved a Mean Squared Error (MSE) of 0.007, actually outperforming the original black-box's error rate on the test set. This suggests that the metamodel successfully captured the underlying signal while ignoring the noise inherent in the original model's training.
Interpretability and Feature Importance
Unlike Meijer G-based methods, SMPF produces closed-form expressions that are easily readable. On the XOR feature selection benchmark, SMPF identified relevant features more accurately than popular local explanation tools like LIME, SHAP, and DeepLIFT. By penalizing overly complex trees during the GP phase, the algorithm naturally favors sparse, elegant solutions, preventing the "formula bloat" common in other symbolic methods.
Computational Efficiency
By utilizing simple primitive functions rather than computationally expensive special functions, SMPF is significantly faster than its predecessors. In comparative tests, SMPF completed its search in approximately 45 minutes, whereas competing symbolic methods required over 215 minutes to reach a solution.
Conclusion
SMPF represents a significant step forward in the field of interpretable machine learning. By combining the theoretical foundations of the Kolmogorov Superposition Theorem with a robust memetic optimization strategy, it provides a way to extract human-readable logic from black-box models without sacrificing accuracy. While the method currently faces challenges with extremely high-dimensional data (like images), it offers a powerful tool for interpreting tabular data in domains where transparency is paramount. Future work may explore deeper tree structures and more sophisticated optimization techniques to further expand the capacity of these symbolic surrogates.
Summary updated