Contrastive knowledge embedding with discriminative self-weighted sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge Graph (KG) embedding maps KG components to low-dimensional spaces.
• Existing KG embedding models focus on scoring functions, neglecting learning frameworks.
• Contrastive Learning (CL) offers potential for representation learning in KG embeddings.

Purpose of the Study:

• To introduce a novel CL framework for KG embedding that addresses the inefficiency of traditional negative sampling.
• To enhance the expressiveness and performance of KG embedding models.

Main Methods:

• Developed a flexible CL framework named "Contrastive knowledge embedding with Discriminative Self-weighted sampling" (CoDiSS).
• Implemented an adaptive weighting mechanism for negative triplets based on their learning contribution.
• Introduced a Discriminative Weight Refinement (DWR) loss to reshape negative score distributions.

Main Results:

• CoDiSS adaptively assigns importance to negative triplets, unlike uniform sampling.
• The DWR loss effectively separates informative from false negatives.
• CoDiSS improves the performance of various KG embedding models (TransE, ComplEx, HousE).

Conclusions:

• The proposed CoDiSS framework enhances KG embedding models by learning from informative negatives and mitigating false negatives.
• CoDiSS leads to more expressive KG embeddings.
• This approach offers a promising direction for advancing KG embedding techniques.