Introduction
As the Internet of Things (IoT) expands, data is increasingly represented as heterogeneous graphs—complex networks where nodes (e.g., devices, sensors, users) and edges (e.g., interactions, locations) possess distinct types and semantic meanings. Analyzing these graphs at scale presents a significant challenge for traditional Graph Neural Networks (GNNs) because the data volume often exceeds available memory.
Existing sampling techniques typically face two major limitations. Random walk-based methods are effective at capturing local neighborhoods but often fail to preserve the global structural integrity of the graph. Conversely, node-wise sampling methods, which select nodes based on simple importance metrics, frequently overlook the complex, multi-hop semantic relationships—known as meta-paths—that define the underlying data logic (e.g., Movie-Actor-Movie).
The HETEROSAMPLE Framework
To address these gaps, we propose HETEROSAMPLE, a deterministic sampling framework designed to create representative subgraphs that preserve both structural and semantic richness. The framework operates in three distinct stages:
- Top-Leader Selection: Instead of stochastic selection, the algorithm identifies influential "leaders" for every node type. It calculates a weighted importance score based on a node's degree and the inherent significance of its type, ensuring the initial sample contains a representative core of the original network.
- Balanced and Guided Expansion: The framework employs Balanced Neighborhood Expansion (BNE) to maintain the original distribution of node types, preventing bias toward dominant entities. Simultaneously, Meta-path Guided Neighborhood Expansion (MGNE) prioritizes neighbors that complete critical local semantic patterns.
- Meta-path Guided Global Sampling: To capture long-range dependencies, HETEROSAMPLE performs guided walks along predefined meta-paths. By prioritizing paths with higher importance weights, the algorithm ensures that semantically relevant multi-hop connections are preserved in the final subgraph.
Experimental Evaluation
We evaluated HETEROSAMPLE on three standard benchmarks: DBLP (bibliographic), ACM (citation), and IMDB (movie database). The method was compared against state-of-the-art baselines, including HGSampling and HetGNN.
Key Results
- Performance Gains: HETEROSAMPLE achieved up to a 15% increase in F1 scores across downstream tasks such as link prediction and node classification.
- Efficiency: The deterministic nature of the algorithm reduced total training runtime by approximately 20% compared to stochastic methods.
- Structural Fidelity: The framework consistently yielded the lowest scores in Node Type Distribution Similarity (NTDS) and Edge Type Distribution Similarity (ETDS), indicating that the sampled subgraphs are mathematically highly representative of the original global structures.
Conclusion
HETEROSAMPLE demonstrates that preserving the specific distribution of node and edge types is as critical as connection density for effective graph representation learning. By utilizing meta-paths as a guide for deterministic sampling, the framework maintains the "semantic soul" of the graph while significantly reducing computational overhead. Future work will focus on extending this approach to dynamic IoT data streams and distributed implementations for even larger-scale networks.
Summary updated