ConNR: A continual N-ary knowledge reasoner for growing N-ary knowledge graphs.

Area of Science:

• Artificial Intelligence
• Data Science
• Knowledge Representation

Background:

• N-ary Knowledge Graphs (NKGs) contain complex, multi-entity facts prevalent in real-world data.
• Existing Link Prediction in NKGs (LPN) methods assume static graphs, requiring costly retraining or fine-tuning when new facts emerge.
• This leads to challenges like knowledge forgetting and significant computational overhead in dynamic environments.

Purpose of the Study:

• Introduce a new task: Continual Link Prediction in growing NKGs (CLPN).
• Address the limitations of static LPN methods in dynamic, evolving knowledge graph scenarios.
• Develop a model capable of adapting to new information without compromising previously learned knowledge.

Main Methods:

• Propose ConNR (Continual N-ary knowledge Reasoner) for CLPN.
• ConNR features an embedding generation module for new elements, an embedding update module for global consistency, and a fact decoder for semantic understanding.
• Utilize five carefully constructed datasets representing diverse NKG growth patterns for evaluation.

Main Results:

• ConNR demonstrates superior performance in predicting missing elements within growing NKGs.
• The model effectively handles the introduction of new facts over time.
• Experimental results validate ConNR's ability to adapt and learn continuously.

Conclusions:

• CLPN is a crucial task for managing and enriching real-world, dynamic NKGs.
• ConNR offers an effective solution for continual link prediction, mitigating knowledge forgetting and reducing computational costs.
• The proposed approach advances the field of knowledge graph completion in evolving data settings.