Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication

Graph alignment—the task of identifying corresponding nodes across multiple graphs—faces major challenges due to oversmoothing in graph neural network embeddings and misaligned latent spaces.
We introduce GADL, a novel framework that integrates a dual-pass spectral encoder with geometry-aware latent space communication.

Our dual-pass encoder combines low-pass and high-pass spectral filters, producing node embeddings that are both structure-aware and discriminative.
To align latent spaces, we incorporate a functional map module that learns bijective and isometric transformations between graph embeddings, ensuring consistent geometric relationships.

Experiments show that GADL:

• Outperforms existing unsupervised alignment baselines on benchmark graphs
• Remains robust under structural and feature inconsistencies
• Generalizes to vision-language alignment, enabling unsupervised multimodal correspondence

[arXiv]

Recommended citation: Maysam Behmanesh, Erkan Turan, and Maks Ovsjanikov. (2025). *Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication.* Preprint, under review.
Download Paper