Incremental Uncertainty-aware Performance Monitoring with Labeling Intervention

Alexander Koebler, Thomas Decker, Ingo Thon, Volker Tresp, Florian Buettner

TL;DR

This paper introduces Incremental Uncertainty-aware Performance Monitoring (IUPM), a novel label-free method to track machine learning model performance during gradual, temporal distribution shifts. IUPM not only estimates performance but also quantifies its uncertainty, enabling an active labeling strategy to efficiently improve accuracy when needed.

What Problem Does This Paper Address?

Machine learning models deployed in the real world often encounter "distribution shifts," where the data they see in production differs from the data they were trained on. These shifts can be gradual and continuous over time, leading to a slow, unnoticed degradation of model performance. Directly monitoring performance is challenging because labeled data is typically unavailable in production, and obtaining labels is costly and time-consuming. Existing methods for estimating performance without labels often fail to quantify the uncertainty of their estimates or effectively leverage the temporal nature of these shifts. This paper tackles the problem of reliably monitoring and understanding model performance degradation in the face of gradual, temporal distribution shifts, especially when labeled data is scarce.

Methods

The core of the proposed method, Incremental Uncertainty-aware Performance Monitoring (IUPM), relies on Optimal Transport (OT). OT is used to find the most cost-efficient way to map data distributions from one time step to the next. Specifically, the authors model the gradual shift by composing a series of incremental OT couplings between consecutive time steps. This allows them to estimate the probability of a label for a new data point by effectively "transporting" information from the initial labeled dataset.

A key innovation is the quantification of uncertainty in these performance estimates. IUPM calculates the standard deviation of the sample-wise loss estimates, providing a measure of how confident the model is in its performance estimation at any given time.

Furthermore, IUPM introduces an active labeling intervention strategy. When the uncertainty in the performance estimate exceeds a predefined threshold, the system identifies the samples that contribute most to this uncertainty. These "critical" samples are then prioritized for labeling, allowing for efficient use of a limited labeling budget to quickly reduce uncertainty and improve the accuracy of the performance estimation.

Datasets

The paper evaluates IUPM on:

  • Synthetic 2D datasets: "Clusters," "Moons," and "Circles" datasets are used to simulate gradual shifts (rotation and translation) in a controlled environment.
  • MNIST dataset: This dataset is used with image perturbations (rotation, translation, scaling) to mimic more complex, real-world shifts.

No new benchmarks were introduced; instead, existing datasets were adapted to simulate temporal distribution shifts.

Key Insights and Ideas

  • Leveraging Temporal Structure: By modeling shifts incrementally using Optimal Transport, IUPM effectively captures the gradual nature of temporal distribution shifts, which is crucial for accurate performance monitoring.
  • Uncertainty is Key: Quantifying the uncertainty of performance estimates is not just an academic exercise; it's essential for building trust and enabling intelligent interventions.
  • Targeted Labeling is Efficient: Prioritizing labeling of samples that contribute most to uncertainty is significantly more effective than random sampling, especially under budget constraints. This makes the system more practical for real-world deployment.
  • Unified Framework: IUPM provides a single framework that can both estimate performance and guide labeling efforts, offering a more holistic approach to model monitoring.

Key Results

  • Superior Performance Estimation: Across various synthetic datasets and models (Random Forest, XGBoost, MLP), IUPM consistently achieved lower Mean Average Error (MAE) in accuracy estimation compared to baseline methods like Average Confidence (AC), Difference Of Confidence (DOC), and Importance Weighting (IM). For instance, on the "Circles" dataset with an XGBoost model, IUPM achieved an MAE of 0.0837, significantly lower than AC's 0.3753.
  • Dramatic Uncertainty Reduction with Intervention: The Uncertainty Intervention (UI) strategy, which selectively labels uncertain samples, drastically reduced the MAE. For example, on the "Circles" dataset with an XGBoost model, IUPMUI achieved an MAE of 0.0242, a nearly 4x improvement over IUPM.
  • Effective on Complex Shifts: On the MNIST dataset with rotational shifts, IUPM demonstrated higher fidelity in performance estimation compared to baselines. The label intervention further improved accuracy and limited uncertainty to a predefined threshold.

Limitations and Future Work

  • Computational Cost of OT: While effective, Optimal Transport can be computationally intensive, especially for high-dimensional data or large sample sizes. The paper mentions using entropic regularization and logarithmic Sinkhorn for efficiency, but this remains a potential bottleneck.
  • Assumptions: The method implicitly assumes that the decision boundary around shifted data points is similar to those linked via the cost-efficient coupling, which might not hold for all types of shifts.
  • Future Work: The authors suggest exploring gradual domain adaptation techniques, active testing, and scaling IUPM to more complex datasets with real-world shifts. Integrating insights from these related fields could further enhance the robustness and applicability of IUPM.

Summary updated

Abstract

We study the problem of monitoring machine learning models under temporal distribution shifts, where circumstances change gradually over time, often leading to unnoticed yet significant declines in accuracy. We propose Incremental Uncertainty-aware Performance Monitoring (IUPM), a novel label-free method that estimates model performance by modeling time-dependent shifts using optimal transport. IUPM also quantifies uncertainty in performance estimates and introduces an active labeling strategy to reduce this uncertainty. We further showcase the benefits of IUPM on different datasets and simulated temporal shifts over existing baselines.
BibTeX
@inproceedings{koebler2025incremental,
  author = {Alexander Koebler and Thomas Decker and Ingo Thon and Volker Tresp and Florian Buettner},
  title = {{Incremental Uncertainty-aware Performance Monitoring with Labeling Intervention}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=eX94LTe7f1},
}
Figure