Continual Test-Time Training on Graphs via Adaptive Prompts Integration

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Continual Test-Time Training (CTTT) on graphs faces challenges with out-of-distribution (OOD) data.
• Existing methods struggle with long-term knowledge retention and efficiency in dynamic graph environments.
• Conventional continual learning requires labeled data and struggles with dynamic OOD graphs.

Purpose of the Study:

• To develop a novel framework for Graph Continual Test-Time Training (GCTTT) that enables continuous adaptation to evolving OOD graphs without supervision.
• To address catastrophic forgetting and improve efficiency in dynamic graph environments.
• To enable a frozen pre-trained graph model to adapt continuously.

Main Methods:

• Proposes Dynamic Prompts-based Continual Graph Learning (DPCGL), a data-centric framework using adaptive prompt optimization.
• Freezes the pre-trained backbone and maintains a dynamic prompt pool for adaptive selection and updating.
• Jointly optimizes similarity alignment, KL divergence regularization, and diversity constraint for stability and adaptability.

Main Results:

• DPCGL achieves state-of-the-art performance on multiple evolving OOD graph benchmarks.
• Effectively alleviates catastrophic forgetting in continual adaptation scenarios.
• Demonstrates robust continual adaptation across domains with parameter-efficient learning.

Conclusions:

• DPCGL offers an effective solution for Graph Continual Test-Time Training in dynamic and OOD environments.
• The prompt-based approach mitigates forgetting by organizing knowledge within prompts.
• The framework provides both stability and adaptability for continuous learning on evolving graphs.