GraphKAN: An Efficient and Interpretable Kolmogorov-Arnold Graph Network for Source Detection

Le Cheng, Peican Zhu, Yangming Guo, Chao Gao, Keke TangOpenReview 2026 · Submission to ICLR 2026 · Withdrawn · OpenReview

Its authors withdrew this paper, so there is no full text to read or summarize.

Abstract

Source detection in graphs offers a viable solution to critical challenges such as rumor tracing. Yet existing GCN-based approaches squander non-embedding parameters and rely on fixed activation functions. We present GraphKAN: An Efficient and Interpretable Kolmogorov–Arnold Graph Network for Source Detection, which capitalizes on Kolmogorov–Arnold Networks (KANs) by assigning learnable activation functions to edge weights. Node features are first diffused through B-spline–based univariate activations, yielding expressive and localized transformations. We further devise a sparsity-aware neighborhood aggregation rooted in community clusters, where edge-level attention is adaptively strengthened through KAN-driven kernel learning. Unlike black-box GCNs, GraphKAN exposes interpretable intermediate representations via its learnable basis functions. Extensive experiments on twelve real-world datasets demonstrate that GraphKAN consistently outperforms state-of-the-art baselines in accuracy, efficiency, and interpretability.
BibTeX
@misc{cheng2026graphkan,
  author = {Le Cheng and Peican Zhu and Yangming Guo and Chao Gao and Keke Tang},
  title = {{GraphKAN: An Efficient and Interpretable Kolmogorov-Arnold Graph Network for Source Detection}},
  year = {2026},
  url = {https://openreview.net/forum?id=qpHDnFUsJo},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}
Figure