A benchmarking framework and dataset for learning to defer in human-AI decision-making

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Machine Learning

Background:

• Learning to Defer (L2D) algorithms are vital for human-AI collaboration in high-stakes domains like fraud detection.
• Current L2D benchmarks often use simplified simulated experts due to the high cost of real expert data.
• This limits the realistic evaluation of L2D systems in critical applications.

Purpose of the Study:

• To introduce OpenL2D, a novel framework for generating synthetic experts with adjustable parameters for L2D system evaluation.
• To create a more realistic benchmark dataset for L2D algorithms using synthetic experts.
• To analyze the impact of expert diversity on L2D algorithm performance.

Main Methods:

• Developed OpenL2D to generate synthetic experts with controllable decision-making processes and work capacity.
• Applied OpenL2D to a public fraud detection dataset to create the Financial Fraud Alert Review (FiFAR) dataset.
• Collected predictions from 50 fraud analysts on 30,000 instances within the FiFAR dataset.
• Evaluated the similarity of synthetic experts to real experts using metrics like consistency and inter-expert agreement.

Main Results:

• The synthetic experts generated by OpenL2D demonstrated comparable consistency and inter-expert agreement to real human experts.
• Performance rankings of different L2D algorithms varied significantly when evaluated with diverse synthetic expert pools.
• The study highlights the critical influence of expert characteristics on L2D algorithm effectiveness.

Conclusions:

• OpenL2D provides a scalable and realistic approach for benchmarking L2D algorithms.
• Realistic expert modeling is essential for accurately assessing L2D system performance in real-world scenarios.
• Future L2D research and development should account for the variability and diversity of human expert behavior.