Accumulated Aggregated D-Optimal Designs for Estimating Main Effects in Black-Box Models

1. Problem Statement

The paper addresses the challenge of main effect estimation in explainable machine learning. The goal is to quantify how individual input variables affect the output of a complex, black-box predictive model ($\hat{f}$) while marginalizing over other variables.

Existing methods suffer from two critical limitations:

1. Sensitivity to Out-of-Distribution (OOD) Evaluations: Methods like Partial Dependence (PD) plots evaluate the model at artificial points that may lie far from the data manifold (especially when features are correlated), leading to unreliable extrapolation.
2. Instability under Feature Correlation: While Accumulated Local Effects (ALE) attempts to mitigate OOD issues by restricting evaluations to local bins, it still relies on a heuristic two-point design (evaluating at the bin boundaries). This design lacks redundancy, resulting in high variance in slope estimation, particularly when features are highly correlated or the surrogate model is sensitive to small perturbations.

Additionally, existing methods like Differential ALE (DALE) require the predictor to be differentiable, excluding non-parametric models like K-Nearest Neighbors (K-NN).

2. Methodology: A2D2E

The authors propose A2D2E (Accumulated Aggregated D-Optimal Designs for Estimating Main Effects), a unified framework that reframes main effect estimation as an experimental design problem.

Core Conceptual Shift

The paper establishes that estimating a main effect function $f_d(t)$ is fundamentally a problem of estimating local slopes (derivatives) and accumulating them.

• Unified View: Existing methods (PD, ALE) differ only in where they choose to evaluate the model to estimate these local slopes.
• The Flaw in Existing Methods: They use a minimal two-point design (e.g., $x_d$ at bin boundaries) which provides no redundancy for variance reduction.

The A2D2E Algorithm

A2D2E replaces the heuristic two-point evaluation with a D-optimal design within a local neighborhood.

1. Local Design Construction:
   For each observation $x_i$ within a specific bin $k$ of the target variable $x_d$, instead of evaluating at two points, A2D2E constructs a local design set $S_{A2D2E, i}$ consisting of the $2p$ vertices of a hypercube centered at $x_i$ with edge length $\delta$.
   $$S_{A2D2E, i} = \left\{ x_i + \frac{\delta}{2}s \mid s \in \{-1, +1\}^p \right\}$$

2. D-Optimality:
   This design maximizes the determinant of the information matrix ($V_i^\top V_i$) for a local linear model. This ensures the minimum variance for the Ordinary Least Squares (OLS) estimator of the local slope $\beta_{k,i}$.

   • Crucially, the symmetry of the hypercube cancels out quadratic terms in the Taylor expansion, providing a second-order accurate approximation of the local slope.

3. Aggregation and Accumulation:

   • Local Slope Estimation: The slope for observation $i$ is computed as a weighted sum of model evaluations at the $2p$ vertices (closed-form solution, no matrix inversion required).
   • Bin Aggregation: These local slopes are averaged within the bin (using the same index set as ALE to avoid OOD issues).
   • Accumulation: The bin-level increments are accumulated across bins to reconstruct the global main effect curve, centered to have a mean of zero.

4. Model Agnosticism:
   The method does not require the model $\hat{f}$ to be differentiable. It only requires the ability to query the model at specific points, making it applicable to K-NN, decision trees, and other non-differentiable black-box models.

3. Key Contributions

1. Unified Formulation: The authors provide a theoretical framework showing that main effect estimation is a design problem, unifying PD and ALE as special cases with suboptimal designs.
2. A2D2E Estimator: Introduction of a new estimator that utilizes D-optimal hypercube designs to minimize the variance of local slope estimation.
3. Theoretical Guarantees:
   • Consistency: Proven that as the number of bins $K \to \infty$ and sample size $n \to \infty$, A2D2E converges to the same population target as ALE.
   • Surrogate Consistency: Extended to show consistency even when only a surrogate model $\hat{f}$ (trained on finite data) is available, provided the surrogate converges uniformly to the true function.
4. Computational Efficiency: While A2D2E evaluates $2p$ points per observation (vs. 2 for ALE), the authors demonstrate that when estimating effects for all dimensions simultaneously, the vertex evaluations can be cached and reused. This results in an amortized cost that is competitive with ALE for moderate dimensions.

4. Empirical Results

The authors conducted extensive simulations across various predictive models (K-NN, Gaussian Processes) and dependence structures (Independent, Low, High correlation).

• Performance under Correlation: A2D2E consistently outperforms PD, ALE, and DALE, with the largest gains observed under high feature correlation.
  • Example: In K-NN regression with high correlation, A2D2E significantly reduced the Overall Root Mean Squared Error (ORMSE) compared to ALE, which suffered from instability.
• Robustness to OOD: In Gaussian Process (GP) experiments with varying length scales (simulating model misspecification), A2D2E remained stable across the entire range, whereas ALE and DALE showed sharp performance degradation at moderate length scales.
• Sensitivity Analysis: The method was found to be robust to the choice of the hypercube radius $\delta$ and the number of bins $K$, provided $K$ is not extremely small.
• Real-World Applications: Demonstrated on the Iris (classification) and Auto (regression) datasets, showing that A2D2E produces interpretable effect curves that align with domain knowledge (e.g., vehicle weight vs. fuel efficiency) and are more stable than PD plots in the presence of correlated features.

5. Significance and Conclusion

This paper makes a significant contribution to the field of Explainable AI (XAI) by bridging the gap between experimental design theory and interpretability.

• Practical Impact: A2D2E offers a robust, model-agnostic alternative to ALE that is particularly vital when dealing with correlated features or non-differentiable models (like K-NN), where current state-of-the-art methods fail or produce high-variance estimates.
• Theoretical Advancement: It moves the field beyond heuristic evaluation strategies by providing a statistically principled approach (D-optimality) to minimize estimation variance.
• Future Directions: The authors suggest extending the framework to handle non-linear local models and heteroscedastic noise, and developing adaptive selection criteria to guide practitioners in choosing between PD, ALE, and A2D2E based on data characteristics.

In summary, A2D2E resolves the trade-off between bias (OOD sensitivity) and variance (estimation instability) in main effect estimation, providing a more reliable tool for understanding complex black-box models.