Multi-relational graph contrastive learning with learnable graph augmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Multi-relational graph learning embeds knowledge graph entities and relations into low-dimensional representations.
• Contrastive learning, with data augmentation, shows promise in addressing sparsity in multi-relational graph learning.

Purpose of the Study:

• To introduce a novel Multi-Relational Graph Contrastive Learning architecture (MRGCL) for improved multi-relational graph learning.
• To enhance performance on multi-relationship prediction tasks by addressing data sparsity.

Main Methods:

• Proposes a Multi-relational Graph Hierarchical Attention Networks (MGHAN) to identify entity importance and extract local graph dependency.
• Learns two adaptive graph augmented views using a variant of MGHAN, adapting to diverse multi-relational graph datasets.
• Designs a subgraph contrastive loss generating positives from strongly connected subgraph embeddings.

Main Results:

• The MRGCL architecture demonstrates superior performance compared to state-of-the-art methods.
• Experiments on multi-relational datasets across three domains validate the effectiveness of the proposed approach.
• The method successfully handles highly sparse data through contrastive learning and data augmentation.

Conclusions:

• MRGCL offers a significant advancement in multi-relational graph learning, particularly for sparse datasets.
• The proposed hierarchical attention and adaptive augmentation strategies are key to its effectiveness.
• The architecture shows broad applicability across various multi-relational graph learning tasks.