On Selective, Mutable and Dialogic XAI: a Review of What Users Say about Different Types of Interactive Explanations

Title of the Paper

On Selective, Mutable and Dialogic XAI: a Review of What Users Say about Different Types of Interactive Explanations

Paper Information

• Topic Area: Interaction and evaluation of human-computer interaction design in Explainable Artificial Intelligence (XAI).
• Keywords: Interactivity, Explainable Artificial Intelligence, Explainability, User Evaluation, Artificial Intelligence, Information Visualization, Human-Computer Interaction

Research Background and Problems

• What issues or challenges did the authors identify?

  • Current XAI technologies predominantly rely on static explanations, with insufficient theoretical and empirical research on interactive explanations.
  • The impact of different types of interactive explanations on user experience (e.g., understanding, utility, and trust) lacks clear conclusions.
  • There is a lack of standardized terminology to describe XAI interaction techniques, making analysis and comparative studies challenging.

• Why is this issue important?

  • The purpose of Explainable Artificial Intelligence is to enhance users' understanding and trust in models, which is critical for practical applications in key fields such as healthcare and law.
  • Interactivity can more effectively support users' cognitive processes for explanations, but its specific advantages and disadvantages remain unclear.
  • Designing more precise explanation systems that meet user needs requires systematic research support.

• Research Motivation and Related Work

  • Inspired by research in human-computer interaction (HCI), information visualization (Infovis), and education, the authors propose that interactive explanations can enhance user understanding and facilitate knowledge construction.
  • The study is based on specific needs summarized from previous theoretical work and experiences, such as Miller's analysis of social science explanation theories and the perspective that XAI systems must meet the needs of diverse user groups.

Solution

• What methods or solutions did the authors propose?

  • Conducted a detailed scoping review, evaluating 48 empirical studies on interactive explanations.
  • Developed an interaction technique taxonomy specific to XAI, categorizing interaction techniques into three levels: "Selective," "Mutable," and "Dialogic."

• What is innovative about this solution?

  • Adapted interaction classification systems from other fields (e.g., information visualization and human-computer interaction) specifically for XAI, proposing nine main interaction types.
  • Systematically summarized the impact of interactive explanations on user experience metrics (e.g., trust, understanding) and identified areas requiring further research.

• What are the implementation steps? What key technologies were used?

  1. Literature Search: Collected XAI-related literature focusing on interactivity and user studies from ACM Digital Library and IEEE Xplore databases.
  2. Taxonomy Development: Extracted explanation types, interaction techniques, and evaluation metrics from the literature, summarizing nine main interaction categories.
  3. Effect Evaluation: Analyzed empirical study data to assess the impact of interactive explanations on user experience metrics.
  4. Result Integration: Proposed an interaction technique taxonomy and user evaluation toolkit, summarizing the advantages and limitations of interactive XAI.

Research Outcomes

• What specific outcomes were achieved?

  • Proposed an interaction taxonomy for XAI, including three levels (Selective, Mutable, and Dialogic) and nine specific interaction types.
  • Found positive effects of interaction mechanisms on user task performance and explanation utility, but the impact on cognitive load and over-reliance issues remains unclear.

• Compared to existing solutions, what are its advantages?

  • Combines theoretical work with empirical research, providing a reference for designing efficient user-centered interactive explanation systems.
  • Systematically summarizes interaction types and evaluation metrics, laying a foundation for future research.

• What are the experimental or evaluation results?

  • Compared to static explanations, interactive explanations significantly improved user task performance and system utility but did not universally enhance usability.
  • The effects of different interaction types may vary. For example, Clarify-type interactions help reduce users' over-reliance on AI, while Simulate-type interactions may increase user workload.

• Limitations and Future Directions

  • Limitations:
    • The study did not conduct a systematic evaluation of the quality of included literature.
    • Primarily based on HCI community literature, potentially missing important research from other AI domains.
    • Current research data mainly comes from laboratory environments, lacking extensive validation in real-world applications.
  • Future Directions:
    • Explore the relationship between interactivity and user memory load to optimize user experience.
    • Investigate how dialogic explanations can be combined with other interaction types to achieve optimal effects.
    • Further clarify the relative contributions of different interaction techniques to improving user understanding and reducing cognitive load.