[Paper Review] Improving Accuracy of Interpretability Measures in Hyperparameter Optimization via Bayesian Algorithm Execution

Julia Moosbauer, Giuseppe Casalicchio+2 more|arXiv (Cornell University)|Jun 11, 2022

Machine Learning and Data Classification4 citations

TL;DR

This paper proposes BOBAX, a Bayesian optimization framework that jointly optimizes for both hyperparameter tuning performance and accurate interpretability via partial dependence plots (PDPs). By integrating Bayesian Algorithm Execution (BAX) to maximize expected information gain (EIG) for IML estimates, BOBAX achieves significantly more reliable PDPs without sacrificing optimization efficiency, enabling trustworthy, real-time interpretability in hyperparameter optimization.

ABSTRACT

Despite all the benefits of automated hyperparameter optimization (HPO), most modern HPO algorithms are black-boxes themselves. This makes it difficult to understand the decision process which leads to the selected configuration, reduces trust in HPO, and thus hinders its broad adoption. Here, we study the combination of HPO with interpretable machine learning (IML) methods such as partial dependence plots. These techniques are more and more used to explain the marginal effect of hyperparameters on the black-box cost function or to quantify the importance of hyperparameters. However, if such methods are naively applied to the experimental data of the HPO process in a post-hoc manner, the underlying sampling bias of the optimizer can distort interpretations. We propose a modified HPO method which efficiently balances the search for the global optimum w.r.t. predictive performance \emph{and} the reliable estimation of IML explanations of an underlying black-box function by coupling Bayesian optimization and Bayesian Algorithm Execution. On benchmark cases of both synthetic objectives and HPO of a neural network, we demonstrate that our method returns more reliable explanations of the underlying black-box without a loss of optimization performance.

Motivation & Objective

• To address the sampling bias in hyperparameter optimization (HPO) that distorts post-hoc interpretability methods like partial dependence plots (PDPs).
• To enable reliable, accurate estimation of marginal hyperparameter effects during the HPO process, rather than relying on post-hoc analysis of biased data.
• To develop a generic, efficient method that balances optimization performance and interpretability quality in automated machine learning.
• To allow practitioners to interact with interpretable results earlier by setting desired precision thresholds for IML estimates.
• To enhance trust in AutoML systems by making the decision process transparent through accurate, real-time explanations.

Proposed method

• Integrates Bayesian Optimization (BO) with Bayesian Algorithm Execution (BAX) to jointly optimize for predictive performance and interpretability.
• Uses Expected Information Gain (EIG) as a surrogate objective to guide sampling toward configurations that improve the accuracy of PDP estimates.
• Computes EIG for PDPs of multiple hyperparameters (e.g., learning rate, dropout, weight decay) in a joint manner to improve global interpretability.
• Employs an adaptive variant (a-BOBAX) that prioritizes interpretability until a user-defined confidence interval threshold (e.g., ±1.5% accuracy) is reached, then reverts to pure optimization.
• Leverages the surrogate model from BO to quantitatively estimate IML statistics, ensuring that interpretation is based on a reliable, uncertainty-aware model.
• Supports any BO variant and can be extended to other IML methods beyond PDPs, such as permutation importance or SHAP.

Experimental results

Research questions

• RQ1 Can we improve the accuracy of interpretability measures like partial dependence plots in hyperparameter optimization without degrading optimization performance?
• RQ2 How can we integrate interpretability objectives directly into the Bayesian optimization loop to reduce sampling bias in IML estimation?
• RQ3 Can we achieve reliable IML estimates earlier in the optimization process by using adaptive confidence thresholds?
• RQ4 Does the proposed method maintain competitive optimization performance compared to standard BO and BAX?
• RQ5 To what extent does the method scale to real-world HPO problems, such as neural network hyperparameter tuning?

Key findings

• a-BOBAX achieved significantly faster convergence to a desired PDP precision threshold (±1.5% accuracy points) compared to standard BO and BAX, enabling earlier reliable interpretation.
• Final PDP estimates from a-BOBAX were as accurate as those from random search (RS), demonstrating high reliability in interpretation quality.
• Optimization performance of a-BOBAX was comparable to that of standard BO with Expected Improvement (EI), maintaining high predictive performance.
• The method reduced misleading interpretations caused by under-sampled regions in the hyperparameter space by proactively balancing exploration for both performance and interpretability.
• Empirical results on 15 diverse datasets showed consistent improvements in PDP accuracy across different HPO scenarios, with strong robustness to dataset-specific characteristics.
• The computational overhead of BOBAX was found to be negligible compared to the cost of training and validating neural network configurations, making it practical for real-world use.