Building Trust in Interactive Machine Learning via User Contributed Interpretable Rules

Title of the Paper

Building Trust in Interactive Machine Learning via User Contributed Interpretable Rules

Paper Information

• Research Area: Interactive Machine Learning, Explainable Artificial Intelligence
• Keywords: Machine Learning, XML, XIML, Tic-Tac-Toe Game, User Studies, User-Centered Evaluation Framework, User Experience

Research Background and Problem

• Identified Issues or Challenges: Traditional machine learning models are often perceived as "black boxes," making it difficult for users to understand the decision-making process. Moreover, while machine learning systems can provide accurate predictions, they often lack the ability to effectively utilize user feedback. This lack of feedback integration may prevent models from correcting errors or incorporating domain-specific logic.

• Importance of the Problem: With the growing application of machine learning technologies, especially in fields requiring high interpretability (e.g., medical diagnostics and security analysis), enhancing user trust and satisfaction has become a critical issue.

• Research Motivation and Related Work: Previous studies have shown that increasing the interpretability and interactivity of machine learning models can enhance user trust. However, the combination of interpretability and interactivity has not been thoroughly explored. Therefore, this study aims to investigate how interactive rules can enhance user trust through a novel system.

Solution

• Proposed Method or Solution: The authors developed an "explanation-driven interactive machine learning system" (XIML) to support user feedback and allow users to edit model rules to improve its performance. The system's use case is based on the Tic-Tac-Toe game, where classification rules are presented in Boolean logic.

• Innovative Contributions:

  1. Combining XML and IML to provide users with an interactive mechanism for understanding and editing rules.
  2. Offering an intuitive grid visualization tool to replace traditional textual rule representations.
  3. Employing a user-centered evaluation framework to comprehensively analyze user experience.

• Implementation Steps and Key Techniques:

  1. Rule Generation: Using the BRCG algorithm (Boolean Rule Column Generation) to generate initial rules trained on the Tic-Tac-Toe dataset.
  2. Interaction Design: Developing a web application that supports user modification of rules, while testing the effects of textual descriptions versus grid-based explanations.
  3. User Experiments: Designing experiments under different conditions (interactive vs. non-interactive; grid visualization vs. textual rules), collecting user feedback, and measuring satisfaction and trust levels.

Research Outcomes

• Specific Findings:

  1. Experiments demonstrated that allowing users to edit model rules significantly improved their perceived control over the system, ultimately enhancing user satisfaction.
  2. The effectiveness of grid-based explanations on user experience varied based on users' educational backgrounds, with high school-educated users preferring textual formats.

• Advantages Over Existing Solutions:

  • While previous studies have separately explored XAI and IML, this research is the first to combine the two and design a mechanism supporting user rule editing.
  • The introduced user interaction features fostered the development of "team cognition" between users and the system, enhancing human-machine collaboration in machine learning.

• Experimental or Evaluation Results:

  • Structural Equation Modeling (SEM) analysis indicated that the system's interactive features had a significantly positive impact, while the effectiveness of grid-based visualization was influenced by individual user characteristics (e.g., educational background).
  • The XIML system was particularly well-received by users familiar with XAI, suggesting that traditional explainable AI plays a crucial role in promoting interactive machine learning.

• Limitations and Future Directions:

  1. Limitations:
     • The testing environment was based on Tic-Tac-Toe, a relatively simple and objective task. Future research should explore more complex tasks.
     • Participant samples were primarily drawn from Amazon Mechanical Turk in the United States, which may lack representativeness.
  2. Future Directions:
     • Testing the XIML system in subjective decision-making domains (e.g., loan approvals).
     • Optimizing interface design for different user groups to ensure both novice and expert users have a positive interactive machine learning experience.

Through the above analysis, this research provides a solid foundation for improving user trust and satisfaction in interactive machine learning while raising several questions worthy of further exploration.