Editable XAI: Toward Bidirectional Human-AI Alignment with Co-Editable Explanations of Interpretable Attributes

Paper Title

Editable XAI: Toward Bidirectional Human-AI Alignment with Co-Editable Explanations of Interpretable Attributes

Publication Info

• Topic area: Explainable AI with editable, rule-based explanations for human-AI alignment.
• Keywords: Editable XAI, human-AI alignment, CoExplain, decision trees, neurosymbolic learning, interactive machine learning, explainable AI, user study, interpretable attributes, bi-directional collaboration.

Background and Problem

• Problem / challenge: Traditional Explainable AI (XAI) systems provide static, read-only explanations, which limit users' ability to correct AI errors or align AI reasoning with their domain knowledge. This creates persistent misalignment and hinders understanding.
• Significance: Addressing this limitation is crucial for improving trust, usability, and alignment in high-stakes domains like medicine and finance, where domain knowledge is critical.
• Motivation and related work: Prior works in XAI focus on one-way communication (e.g., saliency maps, post-hoc justifications) or interactive explanations without allowing users to modify the AI's reasoning. Editable XAI aims to bridge this gap by enabling bi-directional alignment through editable explanations.

Solution

• Proposed approach: CoExplain, a neurosymbolic framework for Editable XAI, allows users to read, write, and enhance rule-based explanations, enabling collaborative human-AI alignment.
• Novelty:
  1. Editable explanations that allow users to directly modify AI reasoning through decision tree rules.
  2. AI-assisted enhancements to user-written rules, including threshold refinement and topology reorganization.
  3. Integration of neurosymbolic methods to align decision trees with neural networks for bi-directional collaboration.
  4. A user interface supporting intuitive rule editing and AI-guided enhancements.
• Procedure and key techniques:
  1. Read: Distill neural network predictions into interpretable decision tree rules.
  2. Write: Parse user-authored decision tree rules into neural network representations.
  3. Enhance: AI refines user rules by optimizing thresholds and topology while preserving user intent.
  4. Regularization techniques ensure alignment between user-defined rules and AI-optimized models.

Results

• Concrete findings:
  • Editable explanations significantly improved user-AI faithfulness (up to 93.9% accuracy in user understanding).
  • CoExplain reduced editing effort by 53% compared to manual editing, requiring fewer operations (9 vs. 14) and iterations (1.59 vs. 3.62).
  • CoExplain achieved near-optimal AI performance (69.9% accuracy on pre-trained data) while maintaining alignment with user rules.
• Advantage over baselines:
  • Editable and CoExplain explanations outperformed read-only explanations in user understanding and alignment.
  • CoExplain balanced alignment and performance better than manual editing, achieving closer alignment to user rules with fewer edits.
• Experiments / evaluation:
  • User study with 43 participants across three tasks (Adult Income, House Price, Heart Disease).
  • Metrics: user-AI faithfulness, AI accuracy, alignment (Tree Edit Distance), editing effort, and perceived ratings.
  • CoExplain was rated highly for ease of use (89.74% acceptance of AI enhancements) and alignment.
• Limitations and future work:
  • Current implementation is limited to structured data and decision tree rules.
  • Future work could extend Editable XAI to unstructured data, larger models, and domain-specific representations (e.g., equations, modular sub-models).

Summary

This paper introduces CoExplain, a neurosymbolic framework for Editable XAI, enabling users to collaboratively align AI reasoning with their domain knowledge through editable decision tree explanations. CoExplain supports reading, writing, and enhancing rules, leveraging neurosymbolic methods to ensure bi-directional alignment. A user study demonstrated that CoExplain improves user understanding, reduces editing effort, and balances alignment with near-optimal AI performance. These findings highlight the potential of editable explanations to foster more effective and efficient human-AI collaboration, with future work aiming to generalize the approach to broader domains and data types.