THeGAU: Type-Aware Heterogeneous Graph Autoencoder and Augmentation

Heterogeneous Graph Neural Networks (HGNNs) are effective for modeling Heterogeneous Information Networks (HINs), which encode complex multi-typed entities and relations. However, HGNNs often suffer from type information loss and structural noise, limiting their representational fidelity and generalization. We propose THeGAU, a model-agnostic framework that combines a type-aware graph autoencoder with guided graph augmentation to improve node classification. THeGAU reconstructs schema-valid edges as an auxiliary task to preserve node-type semantics and introduces a decoder-driven augmentation mechanism to selectively refine noisy structures. This joint design enhances robustness, accuracy, and efficiency while significantly reducing computational overhead. Extensive experiments on three benchmark HIN datasets (IMDB, ACM, and DBLP) demonstrate that THeGAU consistently outperforms existing HGNN methods, achieving state-of-the-art performance across multiple backbones.

💡 Research Summary

The paper “THeGAU: Type-Aware Heterogeneous Graph Autoencoder and Augmentation” addresses two persistent challenges in Heterogeneous Graph Neural Networks (HGNNs): type information loss and structural noise. HGNNs are powerful for modeling complex Heterogeneous Information Networks (HINs) with multiple node and edge types, but they often suffer from oversimplifying type-specific semantics into a unified space and are vulnerable to noisy or missing connections within the graph structure.

To overcome these limitations, the authors propose THeGAU, a novel, model-agnostic framework that synergistically integrates a type-aware graph autoencoder with a guided graph augmentation mechanism. The core idea is to use the reconstruction of schema-valid edges as an auxiliary task to preserve the heterogeneous structural semantics during training, while simultaneously leveraging the trained decoder to refine the graph structure itself.

The THeGAU framework consists of five key components: 1) A Heterogeneous Graph Encoder that can be any existing HGNN backbone (e.g., HGT, SimpleHGN). 2) A main HGNN Classifier (HGC) for the primary node classification task. 3) A Type-aware Graph Decoder (TGD), which is the technical centerpiece. Instead of a shared decoder, TGD employs type-specific MLPs to process node embeddings before using an inner product to predict connections. Critically, it only predicts “legal edges” defined by the graph schema (e.g., Movie-Actor in IMDB), ensuring the model learns to respect the inherent heterogeneity. 4) A Feature-based Classifier (FBC) that connects the initial projection layer of the HGNN directly to a classifier via a skip-layer connection, mitigating over-smoothing and information compression in deep message passing. 5) A Type-aware Graph Augmentation (TG-Aug) module that uses the probabilistic edge predictions from the trained TGD to selectively remove noisy existing edges and add beneficial new ones, thereby denoising and enhancing the input graph.

These components are trained jointly in an end-to-end semi-supervised manner, with a combined loss function balancing the main classification task and the auxiliary tasks of edge reconstruction (via TGD) and skip-layer classification (via FBC). After initial training, the TG-Aug module creates an improved graph, which is then used to retrain the model for final inference.

Extensive experiments on three benchmark HIN datasets (IMDB, ACM, DBLP) demonstrate THeGAU’s effectiveness. When applied on top of four different HGNN backbones, THeGAU consistently and significantly improves their node classification accuracy, often achieving state-of-the-art results. The ablation studies confirm the individual contributions of the TGD auxiliary task and the TG-Aug graph refinement step. The framework’s design also offers practical benefits, such as reduced computational overhead by focusing only on legal edges and using focal loss to handle class imbalance in edge prediction.

In summary, THeGAU presents a comprehensive solution that enhances HGNNs by explicitly preserving type information through a novel autoencoder design and actively improving the input data quality through a decoder-driven augmentation strategy, leading to more robust, accurate, and generalizable models for heterogeneous graph analysis.