[Paper Review] A Study on Mitigating Hard Boundaries of Decision-Tree-based Uncertainty Estimates for AI Models

Pascal Gerber, Lisa Jöckel+1 more|arXiv (Cornell University)|Jan 10, 2022

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

This paper proposes softening the hard decision boundaries in decision-tree-based uncertainty estimation for AI models by integrating alternative ensemble methods into the uncertainty wrapper framework. Using pedestrian detection data from the CARLA simulator and YOLOv3, the study shows that random forests most effectively improve Brier score performance while reducing boundary sensitivity, though at the cost of interpretability and increased runtime complexity.

ABSTRACT

Outcomes of data-driven AI models cannot be assumed to be always correct. To estimate the uncertainty in these outcomes, the uncertainty wrapper framework has been proposed, which considers uncertainties related to model fit, input quality, and scope compliance. Uncertainty wrappers use a decision tree approach to cluster input quality related uncertainties, assigning inputs strictly to distinct uncertainty clusters. Hence, a slight variation in only one feature may lead to a cluster assignment with a significantly different uncertainty. Our objective is to replace this with an approach that mitigates hard decision boundaries of these assignments while preserving interpretability, runtime complexity, and prediction performance. Five approaches were selected as candidates and integrated into the uncertainty wrapper framework. For the evaluation based on the Brier score, datasets for a pedestrian detection use case were generated using the CARLA simulator and YOLOv3. All integrated approaches achieved a softening, i.e., smoothing, of uncertainty estimation. Yet, compared to decision trees, they are not so easy to interpret and have higher runtime complexity. Moreover, some components of the Brier score impaired while others improved. Most promising regarding the Brier score were random forests. In conclusion, softening hard decision tree boundaries appears to be a trade-off decision.

Motivation & Objective

• To address the issue of abrupt uncertainty changes caused by hard decision boundaries in tree-based uncertainty clustering.
• To preserve model interpretability, prediction performance, and computational efficiency while smoothing uncertainty estimates.
• To evaluate alternative methods that replace or augment decision trees in the uncertainty wrapper framework.
• To assess the trade-offs between uncertainty estimation quality, interpretability, and computational cost.

Proposed method

• Integrating five candidate approaches—random forests, k-NN, SVM, logistic regression, and k-means—into the uncertainty wrapper framework to replace or supplement decision trees.
• Using the CARLA simulator to generate synthetic pedestrian detection datasets with controlled input quality variations.
• Applying YOLOv3 as the base object detection model to produce predictions used for uncertainty estimation.
• Employing the Brier score as the primary metric to evaluate uncertainty calibration across different methods.
• Clustering inputs based on feature similarity to assign uncertainty levels, with soft assignment methods replacing discrete tree-based assignments.
• Evaluating the impact of each method on uncertainty smoothing, interpretability, and computational overhead.

Experimental results

Research questions

• RQ1 How do soft assignment methods compare to hard decision trees in reducing uncertainty boundary sensitivity?
• RQ2 Which alternative method provides the best trade-off between uncertainty calibration (Brier score) and interpretability?
• RQ3 What is the impact of each method on runtime complexity and model performance?
• RQ4 To what extent do softening techniques improve uncertainty estimation without degrading prediction accuracy?
• RQ5 How do different components of the Brier score respond to the proposed softening approaches?

Key findings

• All integrated approaches successfully softened the hard decision boundaries of tree-based uncertainty assignments, reducing abrupt uncertainty shifts.
• Random forests achieved the most favorable Brier score improvement among all methods, indicating better uncertainty calibration.
• Despite performance gains, soft methods like random forests and k-NN showed reduced interpretability compared to single decision trees.
• Runtime complexity increased significantly with ensemble-based methods, particularly for random forests and k-NN.
• Some components of the Brier score deteriorated with certain methods, indicating partial trade-offs in calibration quality.
• The study confirms that softening decision boundaries involves a fundamental trade-off between uncertainty smoothness, interpretability, and computational cost.