An interpretable model based on concept and argumentation for tabular data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Interpretability is crucial for AI in high-risk sectors like healthcare and finance.
• Traditional tabular data models lack human-understandable knowledge integration.
• Extracting and utilizing human knowledge for tabular data prediction is challenging.

Purpose of the Study:

• To propose a novel interpretable AI component for tabular data.
• To integrate human-understandable knowledge from data and descriptions into AI models.
• To develop a Concept and Argumentation Model (CAM) for transparent AI decision-making.

Main Methods:

• Developed a quantitative argumentation layer to mine concepts from data and descriptions.
• Constructed a Concept and Argumentation Model (CAM) embedding human-aligned reasoning.
• Implemented quantitative argumentation rules for explicit domain knowledge representation.
• Designed a dialogical explanation method for model interpretability.

Main Results:

• CAM decisions are based on human-level knowledge with intrinsically interpretable reasoning.
• Human-subject evaluations show CAM is comprehensible and explanations are well-accepted.
• Experiments on benchmarks and real-world data demonstrate competitive performance against state-of-the-art models.

Conclusions:

• The proposed Concept and Argumentation Model (CAM) offers an interpretable approach for tabular data.
• CAM effectively integrates domain knowledge, enhancing prediction accuracy and transparency.
• This interpretable AI method shows promise for high-risk applications requiring explainable decisions.