SynHING: Synthetic Heterogeneous Information Network Generation for Graph Learning and Explanation (2401.04133v2)

Abstract: Graph Neural Networks (GNNs) excel in delineating graph structures in diverse domains, including community analysis and recommendation systems. As the interpretation of GNNs becomes increasingly important, the demand for robust baselines and expansive graph datasets is accentuated, particularly in the context of Heterogeneous Information Networks (HIN). Addressing this, we introduce SynHING, a novel framework for Synthetic Heterogeneous Information Network Generation aimed at enhancing graph learning and explanation. SynHING systematically identifies major motifs in a target HIN and employs a bottom-up generation process with intra-cluster and inter-cluster merge modules. This process, supplemented by post-pruning techniques, ensures the synthetic HIN closely mirrors the original graph's structural and statistical properties. Crucially, SynHING provides ground-truth motifs for evaluating GNN explainer models, setting a new standard for explainable, synthetic HIN generation and contributing to the advancement of interpretable machine learning in complex networks.

• The paper presents an automated method that synthesizes heterogeneous networks using motif-based merging to replicate real-world graph structures.
• It demonstrates how configurable merge thresholds and motif counts enable tailored, analytics-friendly datasets for robust GNN explanation studies.
• The work provides benchmark datasets with built-in explanation ground truths, paving the way for fair evaluations of graph neural network interpretability.

Introduction to SynHIN

Graph Neural Networks (GNNs) are powerful tools for machine learning on graph data, which has critical applications ranging from social network analysis to e-commerce fraud detection. One significant challenge in this field is the shortage of public heterogeneous information network (HIN) datasets for testing and improving GNNs, particularly where explainability is vital. Addressing this need, this paper introduces SynHIN, a new method designed to create synthetic HINs that closely mirror the statistical properties of real-world networks.

The Need for Synthetic Data

Real-world HIN datasets are rare and often non-representative, leading to overfitting and bias in GNN models. Research efforts aimed at advancing the interpretability of GNNs are hindered by these limitations and the absence of ground truths for explaining model decisions. Synthetic datasets with inherent explanation capabilities offer a promising solution to this predicament. SynHIN not only generates analytics-friendly synthetic data but also embeds explanation ground truths within the graph, enhancing the interpretability studies for GNNs.

SynHIN's Approach and Contributions

SynHIN identifies frequent motifs — recurring, significant subgraph patterns — in real datasets and employs a novel merge strategy to build clusters with these explanatory motifs, subsequently assembling them into a full synthetic HIN. This technique incorporates In-Cluster Merge and Out-Cluster Merge processes, ensuring the synthetic network's structure and features closely match those of a real-world equivalent.

The paper's main contributions include:

• A new automated methodology to create realistic synthetic HIN datasets, enabling more robust research and testing for explainable AI in the domain of HINs.
• Constructing benchmark datasets with built-in explanation ground truths, propelling the field forward by providing a common platform for the development and assessment of new GNN explanation methods.
• Introducing a modular framework for HIN synthesis, involving steps like motif extraction, subgraph building, merging, pruning, and node feature generation, which can be adapted for diverse dataset requirements.

Experimental Insights and Applications

Experiments conducted by the authors showed how adjusting SynHIN's parameters, such as the merge thresholds and motif counts, could tune the synthetic dataset for different research purposes. The findings endorse SynHIN's utility in generating datasets that enable fair and insightful comparisons of GNN interpretation models. Moreover, the research demonstrates SynHIN's flexibility to create multi-label datasets with rich ground truths that pave the way for more transparent and explainable AI systems.

In conclusion, SynHIN stands as a significant breakthrough for researchers in the field of graph machine learning, particularly in light of its modularity, which allows for customization to suit the specificities of various HIN structures. It enhances the interpretability and generalization capability of AI models, equipping researchers with a reliable tool for benchmarking and overcoming one of GNN's major research barriers— the lack of comprehensive and varied datasets.