Collaborative Human-ML Decision Making Using Experts' Privileged Information Under Uncertainty

Mansoureh Maadi, Hadi Akbarzadeh Khorshidi, Uwe AickelinAAAI 2021 · OpenReview

TL;DR

This paper introduces a novel framework for collaborative decision-making between human experts and Machine Learning (ML) models, specifically designed for high-stakes domains like clinical diagnosis. By incorporating experts' "privileged information" (insights not typically available to ML models during training) and explicitly modeling uncertainty using intervals, their approach aims to improve decision accuracy and reliability.

What Problem Does This Paper Address?

Clinical decision-making is a critical area where human expertise remains indispensable, even with the rise of powerful ML models. This is because human experts excel at handling unstructured data and navigating complex, uncommon situations. However, ML models often struggle with these nuances, and their reliability in critical decisions is frequently questioned. Existing "human-in-the-loop" approaches often position ML models as the primary decision-makers, with humans acting as guides. This paper addresses the gap by proposing a truly collaborative model where both humans and ML models are considered equal decision-makers. Furthermore, it tackles the inherent uncertainty in decision-making by incorporating expert insights that are often overlooked by traditional ML training pipelines.

Methods

The core of this paper is a collaborative decision-making framework that integrates ML models and human experts.

  • Privileged Information: The authors leverage "privileged information" from clinical experts. This isn't just about experts guiding ML training; it's about experts providing their real-time, nuanced estimations (as intervals) during the testing/diagnosis phase, which are typically not recorded and thus unavailable for model training.
  • Uncertainty Modeling with Intervals: To capture the inherent uncertainty in both ML predictions and expert opinions, the paper uses interval modeling.
  • ML Uncertainty: For ML models, uncertainty is captured by generating prediction intervals. This is achieved by using bagging to create multiple versions of a classifier, feeding the same test data to each, and then calculating the interquartile range (IQR) of the resulting class probabilities. This IQR forms the uncertainty interval (UI) for the ML model's prediction.
  • Expert Uncertainty: Clinical experts are asked to provide their opinions as intervals, where the width of the interval reflects their level of confidence or uncertainty.
  • Interval Agreement Approach (IAA): To aggregate the interval-based opinions from multiple ML models and clinical experts, the paper employs the Interval Agreement Approach (IAA). IAA aggregates these intervals into a Type 1 Fuzzy Set (T1 FS), which represents the consensus among all decision-makers.
  • Consensus Decision: The final decision is made by calculating the centroid of the aggregated fuzzy set. A centroid value above a certain threshold (e.g., 0.5) indicates the test sample belongs to the main class.

Datasets

The paper uses a synthetic dataset derived from the Liver Disorders dataset from UCI. This dataset, originally containing 345 samples and 5 features related to alcohol consumption and liver disorders, was adapted for a bi-class classification problem by assigning class 0 to fewer than 3 drinks and class 1 to 3 or more drinks. The authors also generated synthetic data for expert opinions to simulate different scenarios.

Key Insights and Ideas

  • True Collaboration: The framework moves beyond simple human-in-the-loop systems to a genuine collaborative model where both humans and ML models contribute as independent decision-makers.
  • Value of Privileged Information: Expert insights, even if not quantifiable for standard ML training, can significantly enhance decision-making when incorporated at the testing stage.
  • Robustness to Poor Performers: The collaborative model demonstrates robustness, meaning that the poor performance of one group of decision-makers (either ML models or experts) does not drastically degrade the overall decision quality.
  • Uncertainty as a Measure: The width of the intervals serves as a direct measure of uncertainty for individual decision-makers or groups, providing valuable insights into their confidence levels and potentially identifying outliers.
  • Intervals Outperform Point Predictions: The paper suggests that explicitly modeling uncertainty with intervals leads to better decision-making performance compared to traditional point prediction methods that aggregate crisp values.

Key Results

The paper presents results from three experimental scenarios using synthetic data:

  • Scenario 1 (Similar Performance): When both ML models and clinical experts perform similarly, the collaborative model shows comparable or slightly better performance than individual groups. The width of the collaborative model's interval is consistently between those of the individual groups.
  • Scenario 2 (Bad Performance): In cases where one group performs poorly (e.g., clinical experts with very low accuracy), the collaborative model remains robust and its performance is not significantly affected. This highlights the model's resilience.
  • Scenario 3 (Comparison to Point Prediction): The proposed collaborative model, using interval modeling and IAA, consistently outperformed its equivalent point prediction model (using majority voting) in terms of accuracy and F-score. This suggests that capturing uncertainty is crucial for improved decision-making.

Limitations and Future Work

  • Synthetic Data: The experiments were conducted using synthetic data. The authors acknowledge that real-world datasets and scenarios are needed for more accurate analysis.
  • Future Work:
    • Real Data Evaluation: The authors plan to collect real data to further validate the proposed model across different scenarios.
  • Multi-Class Classification: Extending the proposed method to handle multi-class classification problems is a key future direction.

Summary updated

BibTeX
@inproceedings{maadi2021collaborative,
  author = {Mansoureh Maadi and Hadi Akbarzadeh Khorshidi and Uwe Aickelin},
  title = {{Collaborative Human-ML Decision Making Using Experts' Privileged Information Under Uncertainty}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=zsxlDW3TLq},
}
Figure