SynHING: Synthetic Heterogeneous Information Network Generation for Graph Learning and Explanation

ICLR 2026 Conference Submission17692 Authors

Keywords: Synthetic Graph Generation, Heterogeneous Information Networks, Graph Neural Networks, Explainable Artificial Intelligence

Abstract: Graph Neural Networks (GNNs) have achieved remarkable success on relational data, yet their interpretability in heterogeneous information networks (HINs) remains underexplored, largely due to the absence of reliable benchmarks with ground-truth explanations. We introduce \textbf{SynHING}, a synthetic HIN generation framework that supports both graph learning and explainability research. SynHING constructs synthetic graphs by extracting motifs from reference networks, assembling them through motif-guided composition, and refining them via post-pruning to preserve structural and statistical fidelity. Importantly, SynHING is \emph{not limited to the extracted motifs}: users can incorporate their own motifs of interest, which the framework integrates seamlessly into the generated graphs. This flexibility enables controlled and reproducible studies across diverse domains. Experiments on IMDB, Recipe, ACM, and DBLP demonstrate that SynHING produces realistic and semantically consistent HINs, while providing a principled testbed for evaluating Heterogeneous GNNs (HGNNs) and explanation methods. To our knowledge, SynHING is the first framework to enable user-defined, motif-aware HIN synthesis, establishing a foundation for interpretable and reproducible research in heterogeneous graph learning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 17692