Augmented Geometric Multi Resolution Analysis for Graph features representation

NeurIPS 2025 Workshop NeurReps Submission114 Authors

Keywords: geometric representation, graph learning, multi-resolution analysis, wavelets

TL;DR: We extend a multi-scale manifold analysis method for graph dataset and show promising improvements.

Abstract: We present an Augmented Geometric Multi-Resolution Analysis (AGMRA) framework for graph feature representation learning. Our approach extends the classical GMRA framework to graph domains, enabling adaptive multi-scale representation feature for graph structures. We overcome the limitation of original GMRA framework by using dense extraction of wavelet features, which we call AGMRA. By incorporating graph-specific geometric properties and augmenting the analysis with domain-specific features, we demonstrate the effectiveness of our method on real-world network datasets, showing improvements in performance compared to traditional graph representation learning.

Submission Number: 114