heterogeneous graph neural networks

**Heterogeneous Graph Neural Networks (HeteroGNNs)** are **models designed for graphs with multiple types of nodes and edges** — acknowledging that a "User-Click-Item" relation is fundamentally different from a "User-Follow-User" relation. **What Is a HeteroGNN?** - **Input**: A graph where nodes have types (Author, Paper, Venue) and edges have relation types (Writes, Cites, PublishedIn). - **Mechanism**: - **Meta-paths**: specific sequences (Author-Paper-Author = Co-authorship). - **Type-Specific Aggregation**: Use different weights for different edge types (HAN, RGCN). **Why It Matters** - **Knowledge Graphs**: Almost all real-world KGs are heterogeneous. - **E-Commerce**: Users, Items, Shops, Reviews are all different entities. Evaluating them uniformly (Homogeneous) loses semantic meaning. - **Academic Graphs**: Predicting the venue of a paper based on its authors and citations. **Heterogeneous Graph Neural Networks** are **semantic relational learners** — respecting the diverse nature of entities and interactions in complex systems.