[Paper Review] Heat kernel coupling for multiple graph analysis

Michael M. Bronstein, Klaus Glashoff|arXiv (Cornell University)|Dec 11, 2013

Face and Expression Recognition32 references3 citations

TL;DR

This paper introduces heat kernel coupling (HKC), a method for aligning spectral geometry across multiple weighted graphs of different sizes without requiring vertex-wise correspondence. By minimizing modifications to Laplacians so that their heat kernels behave consistently, HKC generalizes Laplacian averaging and enables effective multimodal graph analysis in pattern recognition and manifold learning.

ABSTRACT

In this paper, we introduce heat kernel coupling (HKC) as a method of constructing multimodal spectral geometry on weighted graphs of different size without vertex-wise bijective correspondence. We show that Laplacian averaging can be derived as a limit case of HKC, and demonstrate its applications on several problems from the manifold learning and pattern recognition domain.

Motivation & Objective

• Address the challenge of multimodal graph analysis when no bijective correspondence exists between graphs of different sizes.
• Develop a method to align spectral geometry across multiple graphs using only corresponding functions, not vertex mappings.
• Generalize Laplacian averaging to settings where vertex correspondence is unknown or inapplicable.
• Enable consistent spectral analysis across graphs by coupling heat kernel evolution through minimal Laplacian modifications.

Proposed method

• Define a heat kernel coupling (HKC) problem that minimally modifies Laplacians so that their heat kernels evolve consistently across graphs.
• Formulate HKC as a constrained optimization problem minimizing the Frobenius norm of Laplacian modifications while enforcing heat kernel consistency.
• Use the heat equation solution $ \mathbf{H}_t \mathbf{f} = e^{-t\mathbf{L}} \mathbf{f} $ to align graph signals across different graphs via shared initial functions.
• Derive a closed-form solution for the modified Laplacians using matrix perturbation theory and eigen-decomposition.
• Show that HKC reduces to Laplacian averaging in the limit when bijective correspondence is known.
• Apply HKC to real-world datasets (e.g., NUS-WIDE) using diffusion distances and evaluate performance via precision and mAP metrics.

Experimental results

Research questions

• RQ1 Can we align spectral geometry across multiple graphs without requiring vertex-wise correspondence?
• RQ2 How can we couple heat kernel evolution across graphs with different sizes and topologies?
• RQ3 What is the relationship between HKC and Laplacian averaging in the limit of known correspondence?
• RQ4 Can HKC improve multimodal pattern recognition performance compared to uni-modal or averaged Laplacian methods?
• RQ5 How does HKC perform in real-world retrieval tasks like image tagging and color-based search?

Key findings

• HKC successfully aligns spectral geometry across graphs of different sizes without requiring vertex correspondence, enabling consistent analysis.
• In the NUS-WIDE dataset, HKC improved precision@5 to 87.3% for tags and 83.2% for color modality at $ t = 0.75 $, outperforming uni-modal baselines.
• At $ t = 1.25 $, HKC achieved 80.3% mAP for tags and 74.6% for color, surpassing both uni-modal and averaged Laplacian baselines.
• The method generalizes Laplacian averaging: when correspondence is known, HKC converges to the same solution as averaging.
• Visualizations on ring and manifold graphs show that HKC modifies graphs to align heat diffusion, e.g., by cutting edges to match signal propagation.
• Precision-recall curves demonstrate that HKC-modified diffusion distances (solid lines) outperform both uni-modal and averaged Laplacian baselines (dotted and dash-dot lines).