SKIP: A Prototype-Based Scalable Knowledge Graph Representation Learning Method.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge graph representation learning (KGRL) is crucial for applying models to large real-world knowledge graphs (KGs).
• Anchor-based methods aim to reduce computational costs and parameters in KGRL by encoding entities with a limited set of anchors.
• Existing anchor selection strategies are often basic and can lead to suboptimal KGRL performance.

Purpose of the Study:

• To introduce SKIP, a scalable anchor-based KGRL method that enhances entity anchor selection.
• To leverage prototype information for identifying more representative and effective entity anchors.
• To improve the efficiency and performance of KGRL models on large-scale knowledge graphs.

Main Methods:

• SKIP employs a two-step process: pretraining models to encode entities using topological and textual KG information.
• A prototype learning module (PLM) extracts entity prototypes to guide the sampling of informative entity anchors.
• The method focuses on selecting representative entities that encapsulate valuable prototype information.

Main Results:

• SKIP demonstrates superior performance and effectiveness across various downstream tasks and KG scales.
• On the OGB WikiKG 2 dataset, SKIP achieved comparable performance with a significant reduction in running time (approx. 21.28%) and model parameters (approx. 21.43%).
• The results highlight SKIP's enhanced scalability and efficiency compared to baseline methods.

Conclusions:

• SKIP offers a scalable and effective solution for anchor-based KGRL, outperforming existing methods.
• The prototype-driven anchor selection in SKIP leads to improved performance and reduced resource requirements.
• SKIP presents a promising advancement for applying KGRL to large, real-world knowledge graphs.