GAM Coach: Towards Interactive and User-centered Algorithmic Recourse

Title of the Paper

GAM Coach: Towards Interactive and User-centered Algorithmic Recourse

Paper Information

• Domain: Human-Computer Interaction and Explainable Artificial Intelligence
• Keywords: Algorithmic Recourse, Counterfactual Explanations, Explainability, User-centered, Interactive Systems

Research Background and Problem

• Problems and Challenges:
  • Algorithmic recourse aims to provide actionable recommendations for users affected by machine learning decision systems to change outcomes. However, existing methods assume that model developers understand which input variables users can modify, overlooking the subjectivity of recourse actions and user diversity.
  • Current recourse plans lack feasibility, fail to adequately consider user preferences, and struggle to adapt to users' dynamic needs.
• Significance:
  • With the widespread application of machine learning in high-stakes decisions such as loan approvals, hiring, and university admissions, providing transparent and user-friendly recourse tools can reduce confusion for affected individuals and enhance societal trust in AI technologies.
• Motivation and Related Work:
  • Most current methods generate static "counterfactual examples," which lack flexibility and user involvement.
  • Some studies attempt to improve by grouping or increasing the diversity of recourse options, but limitations remain as user preferences for actionable plans are insufficiently addressed.

Solution

• Method and Innovation:
  • Propose GAM Coach, an interactive, user-centered algorithmic recourse tool that integrates interpretable machine learning models (Generalized Additive Models, GAMs) with linear integer programming to generate customized counterfactual explanations.
  • Design an interactive interface that allows users to explore system behavior and dynamically adjust their recourse preferences, such as the difficulty of modifying features, acceptable ranges, and minimizing the number of variables.
• Implementation Steps and Key Techniques:
  1. Model Selection: Use transparent and structurally simple Explainable Boosting Machines (EBMs) to ensure real-time user interaction.
  2. CF Generation: Employ linear integer programming to generate optimized counterfactual examples, supporting both continuous and categorical features and their interactions.
  3. Interactive Interface:
     • Display diverse recourse plans through the "Coach Menu."
     • Enable users to configure feature preferences, including difficulty and range, via the "Feature Panel."
     • Provide a bookmarking feature to help users save and compare recourse plans.
  4. Open Source and Documentation Support: Develop the tool using modern web technologies, with broad support and an open Python library for researchers.

Research Outcomes

• Specific Results:

  • The tool generates actionable recourse plans that consider user preferences.
  • Users can explore and iteratively refine their recourse options to find satisfactory plans.
  • Experiments validate the effectiveness and usability of GAM Coach.

• Comparison and Advantages:

  • Compared to model-agnostic methods (e.g., genetic algorithms, KD-trees), GAM Coach, based on EBM, generates counterfactual examples that are closer to the original input, sparser, and less prone to failure.
  • The tool is fully transparent, allowing users to adjust and observe model behavior in real time.

• Experimental and Evaluation Results:

  • In experiments involving 41 participants, users found GAM Coach easy to use and helpful in identifying recourse plans that met their needs.
  • Data shows that users tend to prefer counterfactual examples with high actionability and fewer feature changes.
  • Experiments also reveal that transparency enables users to identify potentially counterintuitive decision patterns in the model.

• Limitations and Future Directions:

  • The tool currently relies on transparent models and cannot be directly applied to complex black-box models. Future work could explore simplifying models or optimizing constraints to support more model types.
  • Further research is needed to adapt the tool to real-world scenarios (e.g., loans or government funding) and collaborate with domain experts and actual users.
  • Users may experience information overload or excessive cognitive reasoning, necessitating usability design improvements.
  • Future goals include enhancing the interactive experience from a user-centered perspective and continuing to explore the balance between transparency and algorithmic recourse.