Research map / Graph learning / Graph self-supervised learning

DHAKR: Learning Deep Hierarchical Attention-Based Kernelized Representations for Graph Classification

AAAI 2025 · Graph learning › Graph self-supervised learning · approach: Semi-supervised node representation learning · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DHAKR, an end-to-end differentiable hierarchical attention-based kernel model that learns soft assignments of WL-subtree and shortest-path substructure invariants into composite invariants for graph classification.

Problem

R-convolution graph kernels count isomorphic substructures while ignoring global structure, substructure importance, and the classifier is separate from kernel construction; GNNs pool node features and lose topology.

Approach

Learn a soft assignment matrix to hierarchically merge substructure invariants into composite invariants, apply feature-channel and kernel-channel attention to weight substructures and kernel layers, and feed the multi-level kernel embeddings to a classifier trained end-to-end.

Key results

• DHAKR(WL) reaches 89.87% on MUTAG and 68.76% on PTC(MR), outperforming all listed graph kernels on those datasets.
• DHAKR(SP) reaches 55.41% on Shock versus 37.88% for SPGK, and 76.11% on IMDB-B, the best among compared kernels and deep models.
• Against graph deep learning baselines, DHAKR variants obtain the best average ranks (DHAKR(WL) 1.4, DHAKR(SP) 2.4) in the reported table.

Measured effects reported in the paper

Surprising findings

• Removing feature attention causes large drops (e.g., DHAKR(SP) on MUTAG falls from 88.44% to 80.05%), while removing kernel attention yields only small changes.

Limitations

• Results on PROTEINS are not best overall (DHAKR(SP) is beaten by DHAKR(WL)/others), and the model is limited to WL subtree and shortest-path substructures in the main text.
• Hyperparameter sensitivity: the regularization coefficient gamma has a notable effect, especially for DHAKR(WL).
• Complexity grows with N^2 due to kernel matrices over all sample graphs.

Experimental setup

Datasets: MUTAG PTC(MR) PROTEINS IMDB-BINARY IMDB-MULTI Shock

Benchmarks: MUTAG PTC(MR) PROTEINS IMDB-B IMDB-M Shock

Metrics: classification accuracy average rank

Baselines: GCGK RWGK SPGK WLSK CORE SP CORE WL DGCNN DCNN PATCHY-SAN DGK q-RWNN KerGNN

Builds on: Weisfeiler-Lehman subtree kernel shortest path graph kernel R-convolution kernels channel attention (SENet)

Closest papers by meaning

• Rethinking Kernel Methods for Node Representation Learning on Graphs (NeurIPS 2019 · similarity 0.91)
• Cluster-wise Graph Transformer with Dual-granularity Kernelized Attention (NeurIPS 2024 · similarity 0.87)
• Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.87)
• Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks (NeurIPS 2024 · similarity 0.86)
• Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.86)
• Unsupervised Neighborhood Propagation Kernel Layers for Semi-supervised Node Classification (AAAI 2024 · similarity 0.86)