Interactive Visual Exploration of Rule-Based Model Logic

Area of Science:

• Machine Learning
• Data Visualization
• Human-Computer Interaction

Background:

• Rule-based machine learning models are often considered interpretable.
• Model complexity and rule interdependencies hinder human understanding.
• Extracted rule sets can contain logical inconsistencies despite high accuracy.

Purpose of the Study:

• To introduce a visual analytics methodology for exploring rule-based model logic.
• To support systematic analysis of rule sets and their alignment with domain knowledge.
• To enable detection and refinement of illogical or implausible rules.

Main Methods:

• Integration of overview visualizations, interactive filtering, contradiction analysis, and topic modeling.
• Methodology supports reasoning with and without labeled data.
• Demonstrated through case studies on vessel movement classification and COVID-19 prediction.

Main Results:

• Visual analytics facilitates the detection of illogical or counterintuitive rules.
• The approach aids in assessing the impact of identified inconsistencies.
• Case studies show improved model critique beyond performance metrics.

Conclusions:

• Visual analytics enhances the interpretability and trustworthiness of rule-based models.
• The methodology enables domain-relevant insights through logic-focused critique.
• Supports systematic refinement of complex machine learning models.