Description
In graph classification, traditional methods predominantly pursue fine-grained structural refinements, overlooking the practical potential of a “coarse-to-fine” cascaded strategy that can simultaneously boost both efficiency and accuracy. We propose a multi-kernel graph-classification framework tailored for large-scale graph data that follows a three-step pipeline: graph splitting, kernel specialization, and weight fusion. First,a BFS-based layer-wise decomposition partitions each graph into three subgraph types—star (S), cyclic (C) and tail (T)—which are respectively processed by a direct-product kernel, a Weisfeiler–Lehman subtree kernel and a T-Walk kernel to capture local centrality, community density and propagation-path characteristics. Multi-kernel learning (MKL) is then employed to automatically optimize the weights of the three channel-wise kernel matrices, yielding a fused graph-level representation for classification. Evaluated on a public Chinese social-media rumour-propagation data set containing 3,387 graphs, the proposed method attains 90.11% ± 5.73% accuracy under 5-fold cross-validation, surpassing the best existing direct-product kernel by 5.09 percentage points. Ablation studies confirm the strong coupling and mutual complementarity of the
three subgraph types and their specialized kernels, demonstrating the effectiveness and scalability of the framework on complex graph classification tasks.
| Period | 1 Jan 2025 → 9 Dec 2025 |
|---|---|
| Degree of Recognition | International |