SARF: Aliasing Relation-Assisted Self-Supervised Learning for Few-Shot Relation Reasoning.

Area of Science:

• Artificial Intelligence
• Knowledge Representation and Reasoning

Background:

• Few-shot relation reasoning on knowledge graphs (FS-KGR) is crucial for inferring long-tail relations.
• Current self-supervised learning (SSL) methods for FS-KGR often overlook the relationship between high-frequency and long-tail relations.

Purpose of the Study:

• To propose a novel SSL model, SARF, that leverages aliasing relations (ARs) to enhance FS-KGR.
• To address the limitation of existing SSL methods by utilizing the contextual similarity between relations.

Main Methods:

• Introduced a graph neural network (GNN)-based AR-assist module to encode aliasing relations.
• Developed two fusion strategies (simple summation and learnable fusion) to integrate AR information into the SSL backbone.
• Evaluated the model on three few-shot benchmarks.

Main Results:

• The proposed SARF model achieved state-of-the-art (SOTA) performance on most tested benchmarks.
• Leveraging aliasing relations significantly improved the performance of few-shot relation reasoning.
• The fusion strategies effectively incorporated auxiliary information from ARs.

Conclusions:

• SARF demonstrates the effectiveness of incorporating aliasing relations into self-supervised learning for FS-KGR.
• The model offers a promising approach for improving the inference of under-represented relations in knowledge graphs.
• Future work can explore more sophisticated fusion mechanisms and AR encoding techniques.