Abstract
Graph Convolutional Network (GCN) has achieved extraordinary success in learning effective task-specific representations of nodes in graphs. However, regarding Heterogeneous Information Network (HIN), existing HIN-oriented GCN methods still suffer from two deficiencies: (1) they cannot flexibly explore all possible meta-paths and extract the most useful ones for a target object, which hinders both effectiveness and interpretability; (2) they often need to generate intermediate meta-path based dense graphs, which leads to high computational complexity. To address the above issues, we propose an interpretable and efficient Heterogeneous Graph Convolutional Network (ie-HGCN) to learn the representations of objects in HINs. It is designed as a hierarchical aggregation architecture, i.e., object-level aggregation first, followed by type-level aggregation. The novel architecture can automatically extract useful meta-paths for each object from all possible meta-paths (within a length limit), which brings good model interpretability. It can also reduce the computational cost by avoiding intermediate HIN transformation and neighborhood attention. We provide theoretical analysis about the proposed ie-HGCN in terms of evaluating the usefulness of all possible meta-paths, its connection to the spectral graph convolution on HINs, and its quasi-linear time complexity. Extensive experiments on three real network datasets demonstrate the superiority of ie-HGCN over the state-of-the-art methods.
Introduction
Most real-world data—from academic citations to social networks—exists as Heterogeneous Information Networks (HINs), where nodes (e.g., authors, papers, venues) and edges (e.g., "writes," "published in") have different types. To capture high-order relationships, researchers typically use "meta-paths" (e.g., Author-Paper-Author).
However, existing Graph Convolutional Networks (GCNs) for HINs face significant hurdles. First is efficiency: many models calculate "commuting matrices" to represent meta-paths, leading to dense graphs and $O(N^2)$ complexity, which is prohibitive for large-scale data. Second is interpretability and flexibility: most models require manual meta-path selection or apply global weights to relations, failing to capture "personalized" importance for specific nodes.
The ie-HGCN (Interpretable and Efficient Heterogeneous Graph Convolutional Network) addresses these by automatically discovering the most important meta-paths for specific tasks without massive computational overhead.
Methodology
The ie-HGCN architecture employs a hierarchical aggregation strategy that operates in three primary steps:
- Projection: Since different node types (e.g., "Author" and "Paper") reside in different feature spaces, the model projects them into a common semantic space using type-specific matrices.
- Object-level Aggregation: Instead of constructing complex meta-path graphs, ie-HGCN aggregates information from immediate neighbors of each type separately. This step is mathematically equivalent to spectral graph convolution on bipartite graphs.
- Type-level Aggregation (Personalized Attention): This core innovation uses an attention mechanism to weight each neighbor type. Crucially, this attention is "personalized"—the importance of a neighbor type is calculated based on the target node's own features, allowing the model to adaptively "choose" which meta-paths to follow for each specific object.
By stacking $N$ layers, the model automatically evaluates all possible meta-paths up to length $N$ without explicit calculation.
Key Insights and Results
The model was evaluated on three benchmarks: DBLP (academic), ACM (academic), and IMDB (movies). Key findings include:
- State-of-the-Art Performance: ie-HGCN consistently outperformed ten major baselines (including HAN, GTN, and HGT) in Micro-F1 and Macro-F1 scores.
- Quasi-Linear Scalability: By avoiding dense matrix multiplication, its complexity is $O(N \cdot (|V| + |E|))$, scaling linearly with the number of nodes and edges.
- Automatic Meta-path Discovery: The model "discovers" important paths mathematically by learning weights for neighbor types across layers.
- Theoretical Grounding: The authors proved that their local aggregation is a valid form of spectral convolution.
Conclusion
The ie-HGCN provides a robust framework for learning in HINs by balancing efficiency with interpretability. While it faces challenges like "over-smoothing" at great depths and dependency on initial features, its ability to perform personalized meta-path discovery at scale marks a significant advancement. Future work includes extending this architecture to dynamic networks and tasks like link prediction.
Summary updated