Graph-Powered Interpretable Machine Learning Models for Abnormality Detection in Ego-Things Network.

Area of Science:

• Signal Processing
• Machine Learning (ML)
• Autonomous Systems

Background:

• Interpretability of ML models is crucial for understanding decisions in autonomous systems, especially in high-risk scenarios.
• Abnormality detection is a key function requiring transparent ML model outcomes.
• Existing ML models often lack interpretability, hindering trust and adoption in critical applications.

Purpose of the Study:

• To develop an interpretable machine learning (ML) approach for abnormality detection.
• To enhance the self-awareness (SA) and collective awareness (CA) of agents within autonomous systems.
• To demonstrate the link between incremental model updating and interpretability.

Main Methods:

• Utilized graph matching of semantic vocabulary derived from data and relationships for interpretability.
• Employed data-driven ML techniques for model representation.
• Focused on incremental updating of learned models based on agent experiences.

Main Results:

• Achieved interpretability in abnormality detection through semantic graph matching.
• Demonstrated that incremental learning capabilities are directly related to model interpretability.
• Showcased the approach's applicability in a case study of cooperative vehicle networks (IoT nodes).

Conclusions:

• The proposed ML method enhances interpretability in abnormality detection for autonomous systems.
• Incremental learning and multi-level abstraction are key to achieving interpretable ML models.
• The findings are generalizable to various Internet of Things (IoT) frameworks with diverse agents.