Abstract
A rich class of network models associate each node with a low-dimensional latent coordinate that controls the propensity for connections to form. Models of this type are well established in the literature, where it is typical to assume that the underlying geometry is Euclidean. Recent work has explored the consequences of this choice and has motivated the study of models which rely on non-Euclidean latent geometries, with a primary focus on spherical and hyperbolic geometry. In this paper, we examine to what extent latent features can be inferred from the observable links in the network, considering network models which rely on spherical, hyperbolic and lattice geometries. For each geometry, we describe a latent network model, detail constraints on the latent coordinates which remove the well-known identifiability issues, and present schemes for Bayesian estimation. Thus, we develop a computational procedures to perform inference for network models in which the properties of the underlying geometry play a vital role. Furthermore, we assess the validity of those models with real data applications.
Introduction
In network analysis, nodes are often mapped to a latent space where proximity correlates with the likelihood of a connection. While Euclidean geometry is the standard assumption, real-world networks—characterized by hierarchies or tight clusters—often demand non-Euclidean representations. This paper addresses the challenges of modeling networks in hyperbolic, spherical, and lattice geometries, focusing on two primary hurdles: identifiability and computational scalability.
Methodology
Non-Euclidean Latent Geometries
The choice of geometry serves as a structural hypothesis for the network's underlying organization. The authors explore three distinct spaces:
- Hyperbolic (Poincaré Disk): Ideal for hierarchical, tree-like structures often found in social networks with power-law degree distributions.
- Spherical ($S^d$): Useful for modeling directional data and networks with wrap-around clustering.
- Lattice ($\mathbb{Z}^d$): A discrete geometry used to investigate grid-like constraints.
Generative Model and Inference
The probability of a link $y_{ij}$ between nodes $i$ and $j$ is modeled as a logistic function of their distance $d_{\mathcal{G}}(z_i, z_j)$ in the latent space $\mathcal{G}$. The generative process is formalized in the following algorithm:
To address the identifiability issues inherent in distance-based models (where rotations and reflections preserve likelihood), the authors propose an "anchor coordinate" system. By pinning three specific nodes in a 2D space, they stabilize the latent map, facilitating direct interpretation. For inference, the paper introduces Black Box Variational Inference (BBVI) as a scalable alternative to traditional Markov Chain Monte Carlo (MCMC) methods, treating inference as an optimization problem to reduce computation time.
Experimental Results
The models were evaluated on three benchmark social networks:
- Florentine Family: Marriage ties among elite families, modeled using spherical geometry. BBVI achieved a 3x speedup over MCMC (376s vs 1,227s) with nearly identical results.
- Zachary’s Karate Club: Social ties modeled in hyperbolic space, which successfully captured the hierarchical structure preceding the club's split.
- Sampson’s Monk Dataset: Interpersonal relations analyzed through lattice geometry.
In all cases, the faster BBVI method produced predictive probabilities that clearly distinguished between existing links and non-links, matching the accuracy of the more expensive MCMC approach.
Conclusion
This work demonstrates that non-Euclidean geometries provide a more nuanced lens for network analysis. By solving identifiability issues with anchor coordinates and improving scalability via BBVI, the authors provide a robust framework for uncovering hidden social structures. Future research will explore learning spatial curvature directly from data and further optimizing likelihood approximations for massive datasets.
Summary updated