Scalable Graph Kernels with Continuous Attributes

ICLR 2026 Conference Submission18262 Authors

Keywords: Graph kernels, Continuous node attributes, Scalable graph learning, Explicit graph embeddings

TL;DR: We propose NSPPK, a scalable graph kernel that enriches neighborhood features with path connectors and integrates continuous attributes directly, yielding explicit embeddings that rival or surpass GNNs without training or hyperparameter tuning.

Abstract:

We introduce the Neighborhood Subgraph Pairwise Path Kernel (NSPPK), a scalable
and interpretable graph kernel for attributed graphs. NSPPK compares
neighborhoods connected through unions of shortest paths and directly integrates
continuous node features without discretization. This yields explicit, sparse
embeddings where graph similarities reduce to a single dot product. Feature
extraction scales near-linearly, parallelizes efficiently, and is fully
deterministic. Across six benchmarks with continuous attributes, NSPPK achieves
the best average rank among graph kernels and frequently matches or outperforms
modern GNNs—without any training or hyperparameter tuning. By combining
scalability, interpretability, and expressive power, NSPPK offers a practical
alternative for graph learning in low-data or reproducibility-critical settings.
Its advantage lies in working robustly when data is scarce, yet scaling efficiently
to hundreds of thousands of graphs when data is abundant.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 18262