
Combining network geometry with Bayesian inference
Ariane Lizotte
2025-05-06
Vertices are placed in metric space. The edge cost increases with its length.

The metric space can be physical (e.g. transportation network, brain network) or not.
Graphs obtained from hyperbolic space reproduce many empirically observed properties.
Krioukov, D., et al. Hyperbolic geometry of complex networks. Phys. Rev. E 82, 036106 (2010).
Zuev, K., et al. Emergence of Soft Communities from Geometric Preferential Attachment. Sci. Rep. 5, 9421 (2015).
Krioukov, D., et al. Clustering Implies Geometry in Networks. Phys. Rev. Lett. 116, 208302 (2016).
Each edge \((u, v)\) exists with probability
\(\mathbb{H}^2\) model
\[\begin{align} \mathbb{P}[(u, v) | x_u, x_v, \beta] = \frac{1}{1+\exp\left\{\beta(d_\mathbb{H}(x_u, x_v) - R)\right\}}, \end{align}\]
where \(d_\mathbb{H}\) is the hyperbolic distance, \(\beta\) is the sigmoid sharpness, \(R\) is the maximal radial coordinate, \(x_u, x_v\in \mathbb{H}^2\) are the positions of vertices \(u\) and \(v\) respectively.
Using an approximation for \(d_\mathbb{H}\), this is equivalent to
\(\mathbb{S}^1\) model
\[\begin{align} p_{uv} = \frac{1}{1+\left(\frac{d_\mathbb{S}(\theta_u, \theta_v)}{\mu \kappa_u, \kappa_v}\right)^\beta} \approx \mathbb{P}[(u, v) | x_u, x_v, \beta], \end{align}\]
where \(d_\mathbb{S}\) is the arc length, \(\mu\) is a scaling factor, \(x_u=(r_u, \theta_u)\) is written by its coordinates, \(\kappa_u = \kappa_0 e^{(\hat{R}-r_u)/2}\) is a rescaling of \(r_u\) and \(\kappa_0\) is the minimum degree.
We want to represent a given graph using a hyperbolic embedding.
This amounts to:
\[ \text{Pairs of vertices are} \begin{cases} \text{close} & \text{if connected in graph,} \\ \text{far} & \text{if not connected in graph.} \\ \end{cases} \]
The likelihood of obtaining a graph \(G=(V, E)\) is simply \[\begin{align} \mathbb{P}[G|\theta, \kappa, \beta] = \prod_{(u, v)\in V^2} {p_{uv}}^{a_{uv}} (1-p_{uv})^{1-a_{uv}}, \end{align}\] where \(a_{uv}=1\) if \(u\) and \(v\) are connected and is \(0\) otherwise.
Many algorithms give a maximum likelihood estimator (MLE). This is challenging because of the abundance of local maxima.
Every algorithm
yields a single embedding.
We currently ignore
We address both issues using a Bayesian approach.
The posterior of the Bayesian \(\mathbb{S}^1\) model is \[\begin{align} p(\theta, \kappa, \beta | G) \propto \mathbb{P}[G|\theta, \kappa, \beta] \; p(\beta) \prod_{v \in V} p(\theta_v)p(\kappa_v), \end{align}\] where the priors are
\[\begin{align} \theta_u &\sim \text{Uniform}[-\pi, \pi), \\ \kappa_u &\sim \text{Cauchy}, \quad \kappa_u >\epsilon,\\ \beta &\sim \text{Normal}, \quad \beta>1. \end{align}\]

Hamiltonian Monte Carlo1 (HMC) and random walk don’t sample properly.
Since edges are local, groups of nearby vertices should be moved together.
Cluster transformations + random walk yield good samples of the posterior.
Conflicting ground truth model:
A vertex \(v\) is given two positions \(\theta_v^{(1)}\) and \(\theta_v^{(2)}\).
When generating \(G\) with the \(\mathbb{S}^1\) model, each edge probability including \(v\) uses randomly \(\theta_v^{(1)}\)or \(\theta_v^{(2)}\).
Paper:
Lizotte, A., Young, J.-G. and Allard A. Symmetry-driven embedding of networks in hyperbolic space. arXiv:2406.10711 (2024).
[Accepted at Communication Physics]
Antoine Allard
Jean-Gabriel Young
Go to the neighbour closest to the destination.
The \(\mathbb{S}^1\) model is not identifiable because of graph and space symmetries.
Comparing embeddings requires alignment.