Physics-Informed Explainable Continual Learning on Graphs.

Area of Science:

• Artificial intelligence
• Graph learning
• Scientific domains (chemistry, biomedicine)

Background:

• Temporal graph learning models are often black boxes, lacking explainability.
• Understanding information evolution in dynamic graphs is crucial for AI applications in science.
• Existing methods struggle to transparently explain model adaptation to new data.

Purpose of the Study:

• To develop a novel, explainable continual learning method for temporal graphs.
• To enhance transparency and trustworthiness of AI models in scientific applications.
• To address the challenge of explaining how temporal graph learning models adapt to evolving information.

Main Methods:

• Proposes Physics-Informed Explainable Continual Learning (PiECL) for temporal graphs.
• Utilizes physical and mathematical algorithms to quantify data disturbance and detect changes.
• Leverages physics-based theories for a transparent learning mechanism.

Main Results:

• PiECL successfully explains the learning process in temporal graph models.
• The proposed method demonstrates superior performance over state-of-the-art techniques.
• Experimental validation on three real-world datasets confirms effectiveness.

Conclusions:

• PiECL offers a transparent approach to understanding information evolution in temporal graphs.
• The method significantly enhances the explainability of AI models in scientific contexts.
• PiECL holds substantial potential for advancing AI applications in chemistry and biomedicine.