CRII:III: Towards a Continual Collaborative Framework for Knowledge Graph Reasoning

About This Grant

Knowledge graphs (KGs) have become essential tools for organizing and representing complex real-world knowledge. Effectively reasoning over knowledge graphs to extract meaningful insights and uncover novel connections, is crucial for addressing multifaceted national challenges and facilitating interdisciplinary collaboration. However, current KG reasoning methods face significant challenges related to reliability, sustainability, and extensibility. Reliability issues arise due to varying amounts and qualities of available data across different KGs, resulting in inconsistent reasoning outcomes. Sustainability is compromised by the continual evolution of knowledge graphs, which traditionally requires costly and inefficient model retraining whenever new information emerges. Additionally, extensibility remains a critical limitation, as existing reasoning models cannot easily or rapidly adapt to newly developed knowledge graphs, requiring extensive computational resources and hindering timely deployment. This project seeks to overcome these challenges by creating a continual collaborative framework specifically designed for practical, real-world scenarios. The proposed framework will enable reasoning models to incrementally integrate new information, collaboratively share insights across multiple domains while ensuring data privacy and quickly adapt to new KGs. By addressing the critical issues, the project will contribute to the advancement of open knowledge networks, promoting their development in a manner that ensures reliability, sustainability, and extensibility. These improvements will strengthen existing reasoning approaches and enhance their ability to serve as a vital component of the nation's infrastructure. The continual collaborative framework faces two significant challenges: the challenge of catastrophic forgetting and the challenge of effective and robust knowledge transfer. To overcome these challenges, the project proposes two thrusts. The first thrust aims to enable continual local updating of knowledge graph reasoning models. The team will address the challenge of catastrophic forgetting when continually updating KG reasoning models by preserving a set of representative data for rehearsal during subsequent continual updates. For novel emerging data, the team will design a synergistic solution combining prompt tuning and model modulation to tackle knowledge collision and facilitate knowledge integration. The second thrust aims to build an extensible collaborative framework for knowledge graph reasoning. The team will target the challenge of effective and robust knowledge transfer in a collaborative environment where multiple KG reasoning models share knowledge through prompt consolidation and inter-model transfer. To customize reasoning model parameters for emerging KGs, a conditional generative framework will be employed to learn the distribution of parameters from existing reasoning models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.