Self-supervised heterogeneous hypergraph learning with context-aware pooling for graph-level classification

Representation learning in unlabeled heterogeneous graphs has gained significant interest. The heterogeneity in graphs not only provides rich information but also poses challenges to model complex relations in self-supervised learning (SSL) manner. Existing SSL-based approaches are usually designed for node-level tasks and are unable to capture global graph-level features. Also, they often employ computationally expensive meta-path-based techniques, to learn the intrinsic graph structure, that are intractable. Importantly, they overlook non-pairwise relationships among nodes in heterogeneous graphs, for instance in protein-protein interaction networks or collaboration networks, limiting the effectiveness of graph-level learning. To address these issues, we propose a novel self-supervised heterogeneous hypergraph learning framework that captures the richness of heterogeneity, and high-order connectivity in graph-level classification. Unlike traditional methods that rely on meta-path-based approaches to incorporate high-order information, we introduce a k-hop neighborhood strategy to construct intra-graph hyperedges, and a shared attribute-based approach for inter-graph hyperedges to construct the heterogeneous hypergraph. Furthermore, we introduce a context-aware graph-level pooling mechanism that facilitates adaptive aggregation of relevant information across the hypergraph, considering both local and global contexts. Lastly, we design a self-supervised contrastive learning framework by introducing a high-order-aware adaptive augmentation mechanism. This enables the model to learn meaningful graph-level representations from less-labeled data. We evaluate our proposed model against graph kernels, graph neural networks, and graph pooling-based baselines on real-world datasets, demonstrating an overall performance improvement of 5.81% that validates the effectiveness and superiority of the proposed method.