Inductive Meta-Path Learning for Schema-Complex Heterogeneous Information Networks.

Area of Science:

• Computer Science
• Data Mining
• Artificial Intelligence

Background:

• Heterogeneous Information Networks (HINs) feature diverse node and edge types.
• Meta-paths offer semantic explanations but face scalability issues in schema-complex HINs like knowledge bases.
• Exhaustive meta-path enumeration is computationally intensive for large-scale HINs.

Purpose of the Study:

• To introduce SchemaWalk, an inductive meta-path learning framework for schema-complex HINs.
• To address the computational challenges of meta-path enumeration and assessment in large HINs.
• To develop a method for learning meta-path scores without enumerating all path instances.

Main Methods:

• SchemaWalk utilizes schema-level representations for meta-paths.
• A reinforcement learning-based agent navigates the network schema to discover meta-paths.
• The framework learns policies for high-coverage and high-confidence meta-paths across multiple relations.

Main Results:

• SchemaWalk effectively learns meta-paths in schema-complex HINs.
• The proposed approach mitigates the need for exhaustive path instance enumeration.
• Experiments on real datasets validate the framework's effectiveness.

Conclusions:

• SchemaWalk provides an efficient solution for meta-path learning in complex HINs.
• The framework enhances the explainability and utility of meta-paths in knowledge bases.
• This inductive approach offers a scalable alternative to traditional meta-path methods.