TL;DR
This paper introduces LGKDE, a learnable kernel density estimation framework for graphs that leverages graph neural networks and maximum mean discrepancy to improve density estimation and anomaly detection, with theoretical guarantees and superior empirical performance.
Contribution
LGKDE is the first method to learn graph kernel parameters via neural networks for multi-scale KDE, enhancing graph density estimation and anomaly detection.
Findings
LGKDE accurately recovers synthetic graph densities.
LGKDE outperforms baselines in graph anomaly detection.
Theoretical guarantees ensure robustness and convergence.
Abstract
This work proposes a framework LGKDE that learns kernel density estimation for graphs. The key challenge in graph density estimation lies in effectively capturing both structural patterns and semantic variations while maintaining theoretical guarantees. Combining graph kernels and kernel density estimation (KDE) is a standard approach to graph density estimation, but has unsatisfactory performance due to the handcrafted and fixed features of kernels. Our method LGKDE leverages graph neural networks to represent each graph as a discrete distribution and utilizes maximum mean discrepancy to learn the graph metric for multi-scale KDE, where all parameters are learned by maximizing the density of graphs relative to the density of their well-designed perturbed counterparts. The perturbations are conducted on both node features and graph spectra, which helps better characterize the boundary…
Peer Reviews
Decision·Submitted to ICLR 2026
The paper tackles an important and underexplored problem which is relevant to applications such as anomaly detection and molecular graph analysis. The proposed approach is described clearly and supported by theoretical derivations. The experimental section is comprehensive, covering a wide range of datasets and showing that LGKDE consistently improves upon baseline methods. The presentation is coherent, and the proposed framework bridges traditional kernel-based methods and modern deep learning–
The novelty of the contribution is limited, as the method mainly combines known components (GNN embeddings, MMD distances, and KDE) rather than introducing a fundamentally new concept. The proposed perturbation strategy and contrastive density objective are incremental variations on existing ideas in self-supervised learning and graph anomaly detection. The related work discussion is incomplete; several recent graph density and contrastive learning approaches are not adequately compared or discu
1. The paper is clearly written, with a clear problem statement and a logically organized presentation. 2. The paper focuses on a relatively unexplored but important subarea of nonparametric graph density estimation, which directly underpins graph-level anomaly detection. 3. Synthetic validations recover known distributions, and broad benchmarks for anomaly detection show consistent AUROC, AUPRC, and FPR95 improvements over competitive baselines.
1. While complexity analysis is given, the framework requires pairwise deep MMD computation for KDE and generation of multiple perturbed samples, which might be costly for very large datasets. Empirical runtime/memory profiles for large-scale sparse graphs are lacking. 2. Perturbed samples are not true anomalies. The performance gain depends on the quality of the perturbations, and it is unclear how LGKDE would perform when the perturbations poorly reflect anomalous structures.
1. The paper formulates graph-level density estimation through a theoretically grounded framework that integrates graph neural network embeddings with learnable multi-scale kernel density estimation in an MMD-based space. 2. The authors provide L1-consistency and convergence rate results (Theorems 4.1 and 4.2), establishing statistical soundness and connecting the method to nonparametric theory under intrinsic dimension assumptions. 3. The framework includes structure- and spectrum-aware perturb
1. The distinction between LGKDE and deep density estimation methods is not sharply articulated, leaving unclear where LGKDE provides a fundamental advantage. 2. It is not clear how the learned MMD metric is constrained to prevent overfitting of the density landscape (e.g., through regularization of kernel parameters or Rademacher-style control), and how this constraint is reflected in the stated generalization bound. 3. More remarks or further insights are needed to help the readers better unde
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
