<ul><li>A framework LGKDE is proposed for learning kernel density estimation for graphs.</li><li>LGKDE leverages graph neural networks and maximum mean discrepancy to learn the graph metric for multi-scale KDE with learned parameters.</li><li>The method shows consistency and convergence guarantees, including bounds on error, robustness, and complexity.</li><li>Empirical evaluation demonstrates superior performance of LGKDE in recovering synthetic graph distributions and graph anomaly detection on benchmark datasets.</li></ul>

Learnable Kernel Density Estimation for Graphs

Discover more