Graph transformer for link prediction on N-ary facts

Area of Science:

• Artificial Intelligence
• Data Science
• Graph Theory

Background:

• Hyper-relational knowledge graphs extend traditional KGs with multi-dimensional auxiliary information, increasing representational complexity.
• N-ary Fact Link Prediction faces challenges due to the complex and varied expression forms of N-ary facts compared to binary relations.

Purpose of the Study:

• To address the insufficient utilization of heterogeneous graph structure information in existing N-ary fact representation methods.
• To propose an N-ary graph Transformer (NAGT) model for enhanced N-ary Fact Link Prediction.

Main Methods:

• Development of an N-ary graph Transformer (NAGT) model.
• Incorporation of a novel attention mechanism based on N-ary structural bias.
• Improving the representation of N-ary heterogeneous graphs.

Main Results:

• The NAGT model demonstrates superior performance in extracting structural information compared to existing methods.
• Experimental validation on JF17K, Wikipeople, and WD50K datasets confirms NAGT's effectiveness.
• The model accurately identifies key associations in recommendation scenarios.

Conclusions:

• The proposed NAGT model effectively completes knowledge graphs and shows efficiency and robustness in N-ary Fact Link Prediction tasks.
• NAGT enhances the representation of N-ary heterogeneous graphs, leading to improved prediction accuracy.