CAKGE: Context-Aware Adaptive Learning for Dynamic Knowledge Graph Embeddings.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Knowledge graph embeddings (KGE) are crucial for data representation but struggle with evolving real-world facts.
• Existing KGE models face challenges like catastrophic forgetting or require extensive retraining for new information.

Purpose of the Study:

• To develop a novel model, CAKGE, for dynamic knowledge graph embedding that efficiently integrates new information.
• To address the limitations of existing KGE models in transductive, inductive, and continual learning settings.

Main Methods:

• CAKGE identifies semantic-relevant entities and relational paths using a context-aware fusion module.
• An adaptive message aggregation module with knowledge replay integrates new and existing knowledge without retraining.
• Fused Gromov-Wasserstein distance is used for graph matching to align old and new knowledge semantically and topologically, mitigating catastrophic forgetting.

Main Results:

• CAKGE demonstrates state-of-the-art performance in dynamic KGE modeling.
• The model effectively integrates new knowledge while preserving existing information.
• Theoretical guarantees are provided for CAKGE's expressiveness and reasoning capabilities.

Conclusions:

• CAKGE offers a unified framework for transductive, inductive, and continual learning settings in KGE.
• The proposed model efficiently handles evolving knowledge graphs, overcoming catastrophic forgetting and reducing retraining costs.