Enhancing Feature Selection and Interpretability in AI Regression Tasks Through Feature Attribution

Alexander Hinterleitner, Thomas Bartz-Beielstein, Richard Schulz, Sebastian Spengler, Thomas Winter, Christoph LeitenmeierarXiv 2024 · OpenReview

Abstract

Research in Explainable Artificial Intelligence (XAI) is increasing, aiming to make deep learning models more transparent. Most XAI methods focus on justifying the decisions made by Artificial Intelligence (AI) systems in security-relevant applications. However, relatively little attention has been given to using these methods to improve the performance and robustness of deep learning algorithms. Additionally, much of the existing XAI work primarily addresses classification problems. In this study, we investigate the potential of feature attribution methods to filter out uninformative features in input data for regression problems, thereby improving the accuracy and stability of predictions. We introduce a feature selection pipeline that combines Integrated Gradients with k-means clustering to select an optimal set of variables from the initial data space. To validate the effectiveness of this approach, we apply it to a real-world industrial problem - blade vibration analysis in the development process of turbo machinery.

Introduction

Deep learning models are often characterized as "black boxes" due to their complex, non-linear internal structures. While they excel as universal function approximators, their lack of transparency is a significant hurdle in high-stakes engineering fields like turbomachinery. In these domains, understanding the physical drivers behind a prediction—such as the vibration level of a turbine blade—is as critical as the prediction itself. This paper explores how Explainable AI (XAI) can be transitioned from a tool for legal or ethical justification ("BLUE" XAI) to a tool for engineering improvement: Research, Exploration, and Debugging ("RED" XAI).

The Challenge of Feature Selection in Regression

Most XAI research has historically focused on classification tasks (e.g., identifying objects in images). Regression tasks, which involve predicting continuous values, present unique challenges for interpretability. Standard feature selection methods often fall short in these scenarios. "Filter" methods like Pearson correlation may miss complex non-linear relationships, while "Wrapper" methods that iteratively test feature subsets are often too computationally expensive for deep neural networks.

Comparison of feature selection and attribution methods

As shown in the comparison above, feature attribution methods like Integrated Gradients offer a middle ground, providing theoretically sound insights into model decisions while remaining applicable to complex differentiable models.

Methodology: The XAI-Driven Pipeline

The authors propose a four-stage data-driven pipeline designed to refine feature sets for regression:

  1. Hyperparameter Optimization: Using the Sequential Parameter Optimization Toolbox (SPOT), the authors first ensure the Multi-Layer Perceptron (MLP) is optimally tuned. This is crucial because XAI attributions are only as reliable as the underlying model.
  2. Feature Attribution via Integrated Gradients (IG): IG is employed to attribute the model's predictions to specific input features. It satisfies the axioms of sensitivity and completeness, ensuring that every change in output is accounted for by the input features.
  3. Global Aggregation & Clustering: To move from local (per-sample) explanations to a global understanding, the mean attribution across the dataset is calculated. Features are then grouped into "importance classes" using k-means clustering.
  4. Iterative Pruning: The cluster representing the least important features is removed, and the model is re-tuned to evaluate performance on the streamlined feature set.

Industrial Application: Turbomachinery Blade Vibration

The pipeline was validated using a real-world industrial dataset provided by MAN Energy Solutions. This dataset contains over 27,000 samples aimed at predicting the vibration amplitudes of blades in turbochargers—a critical factor in preventing mechanical failure.

Turbocharger components for blade vibration analysis

The data includes 86 features covering thermodynamic variables and geometric configurations. The XAI analysis identified "Oscillation Mode" and "Excitation Order" as the most critical predictors, a finding that aligns perfectly with established mechanical engineering theory and serves as a vital sanity check for the model's logic.

Key Results and Insights

The results demonstrate that "less is more" when it comes to feature density in complex regression:

  • Accuracy Gains: Reducing the feature set from 86 to the 64 most important variables dropped the Mean Squared Error (MSE) from 8.718 to 5.762. This suggests that the original model was likely hindered by uninformative noise.
  • Enhanced Stability: The standard deviation of the MSE across cross-validation folds decreased from 1.594 to 0.425, indicating a significantly more robust and reliable model.
  • Superiority over Baselines: The IG + k-means approach outperformed traditional methods like Pearson Correlation (MSE 13.026), Lasso Regression (MSE 8.443), and even other XAI methods like KernelSHAP (MSE 7.824).

Conclusion

This research highlights the power of using XAI as a sophisticated filter for feature selection. By leveraging the non-linear patterns captured by neural networks, engineers can identify the most physically relevant features, leading to models that are not only more accurate but also more stable and interpretable. Future work will look toward applying these techniques to even more complex architectures, such as CNNs and LSTMs, to further bridge the gap between deep learning and physical intuition.

Summary updated

Abstract

Research in Explainable Artificial Intelligence (XAI) is increasing, aiming to make deep learning models more transparent. Most XAI methods focus on justifying the decisions made by Artificial Intelligence (AI) systems in security-relevant applications. However, relatively little attention has been given to using these methods to improve the performance and robustness of deep learning algorithms. Additionally, much of the existing XAI work primarily addresses classification problems. In this study, we investigate the potential of feature attribution methods to filter out uninformative features in input data for regression problems, thereby improving the accuracy and stability of predictions. We introduce a feature selection pipeline that combines Integrated Gradients with k-means clustering to select an optimal set of variables from the initial data space. To validate the effectiveness of this approach, we apply it to a real-world industrial problem - blade vibration analysis in the development process of turbo machinery.
BibTeX
@misc{hinterleitner2024enhancing,
  author = {Alexander Hinterleitner and Thomas Bartz-Beielstein and Richard Schulz and Sebastian Spengler and Thomas Winter and Christoph Leitenmeier},
  title = {{Enhancing Feature Selection and Interpretability in AI Regression Tasks Through Feature Attribution}},
  year = {2024},
  eprint = {2409.16787},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.16787},
}
Figure