New Graph Learning Model Outperforms Baselines Across Classification and Clustering Tasks

A new heterogeneous graph structure learning model named HGSL-SPSE has demonstrated superior performance across node classification and unsupervised clustering benchmarks compared to eight mainstream baseline architectures.

The core objective of heterogeneous graph structure learning involves constructing high-quality topological structures, as the precision of structural modeling directly dictates the success of subsequent graph representation learning and downstream analytical tasks. In existing methodologies, systems predominantly depend on raw node features or shallow embeddings to compute similarities for edge pruning and topological rewiring across networks.

However, these conventional techniques remain vulnerable to feature noise and distribution shifts across distinct node types. Consequently, traditional frameworks frequently suffer from the erroneous removal of critical semantic edges, the unwarranted retention of redundant noisy edges, and an ongoing inability to capture complex, higher-order semantic associations that span multi-typed node structures.

To overcome these structural hurdles, the research team designed HGSL-SPSE around a similarity perception mechanism and collaborative structure enhancement framework. The system first deploys a Personalized PageRank-based similarity perception mechanism that deeply fuses raw node features with global topological information. When paired with semantic meta-paths, this mechanism constructs multi-typed structure–semantic association matrices capable of characterizing high-order semantic dependencies among diverse nodes.

Following association modeling, a multi-perspective feature-structure collaborative enhancement module bidirectionally strengthens both the raw feature graph and the structure–semantic graph. An adaptive deep structure optimization module then refines the topology through targeted noise-edge elimination and critical-edge completion, after which the refined topology and reinforced features are processed by a representation learning encoder to yield discriminative node embeddings.

The researchers verified the effectiveness of HGSL-SPSE through empirical evaluations on four public heterogeneous graph benchmarks. Across both node classification and unsupervised clustering evaluations, the architecture consistently surpassed eight mainstream baselines, with the authors releasing their code implementation in a public repository.

Subject of Research: Heterogeneous graph structure learning via similarity perception and structure enhancement

Article Title: Heterogeneous graph structure learning via similarity perception and structure enhancement

Article References: Li, C., Wan, L., Li, X., Liu, R., Gao, J., & Jiang, C. (2026). Heterogeneous graph structure learning via similarity perception and structure enhancement. Applied Intelligence, 56(16), Article 556. https://doi.org/10.1007/s10489-026-07589-0

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07589-0

Keywords: heterogeneous graphs, graph structure learning, similarity perception, structure enhancement, Personalized PageRank, topological modeling, graph representation learning, node classification, unsupervised clustering, semantic meta-paths, structure-semantic association matrices, feature-structure collaborative enhancement, noise-edge elimination, critical-edge completion, discriminative node embeddings, cross-type distribution shifts, HGSL-SPSE, benchmark datasets