Reproduce “Learnable Kernel Density Estimation for Graphs and Its Application to Graph-Level Anomaly Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xudong Wang, Ziheng Sun, Chris Ding, Jicong Fan
LGKDE introduces a framework for learnable kernel density estimation on graphs, integrating deep graph representation learning with adaptive multi-scale KDE to effectively model graph distributions. The method achieves state-of-the-art performance in graph-level anomaly detection, surpassing baselines with an average AUROC of 79.37% across twelve benchmark datasets.
Key insights
- Combining a learned MMD metric space with adaptive multi-scale KDE is crucial for effectively capturing the complex, multi-scale distributions of graph-structured data.
- The unsupervised density contrasting objective, enabled by carefully designed structure-aware graph perturbations, provides an effective signal for learning graph densities without explicit anomaly labels, preventing model collapse.
- Providing theoretical guarantees for consistency, minimax-optimal convergence rates, and robustness significantly enhances the understanding and reliability of deep density estimation models for graphs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.