📊 statistics

Latent space models for networks with nodal multiplicative effects

This paper introduces a generalized latent space model for networks that incorporates nodal multiplicative effects to capture structural heterogeneity through local metric deformations, demonstrating via simulations and real-world applications that this approach enhances generative flexibility and topological accuracy compared to classical models.

Original authors: Carlos Nosa, Juan Sosa

Published 2026-08-10

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Carlos Nosa, Juan Sosa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a giant, invisible web of friendships, rivalries, or alliances. Maybe it's the social circle of a high school, the connections between neurons in a brain, or the trade routes between ancient cities. Scientists call these webs "networks." To make sense of them, researchers often use a clever trick called a latent space model. Think of this as a magical map where every person (or node) is hidden as a dot in a geometric space. The closer two dots are on this map, the more likely they are to be friends or connected. If they are far apart, they probably don't know each other.

For a long time, scientists assumed this map was made of standard, rigid geometry—like a flat sheet of paper (Euclidean), a perfect ball (Spherical), or a weird, saddle-shaped surface (Hyperbolic). They believed that the "distance" between any two people was measured by the same ruler everywhere. If you moved a step to the left, it cost the same amount of "distance" whether you were in the center of the city or the edge of town. But real life is messy. Some people are super popular and seem to be everywhere at once, while others are isolated even if they are physically close to others. The old maps couldn't explain why some dots were "closer" to everyone else without actually moving them on the map. This paper asks: What if the ruler itself changes size depending on who is holding it?

The Paper's Big Idea: Stretching the Ruler

In this study, Carlos Nosa and Juan Sosa propose a fun twist on those old maps. They suggest that instead of using one rigid ruler for the whole network, we should let each person carry their own special ruler that can stretch or shrink. They call this a "nodal multiplicative effect."

Imagine you are playing a game of tag in a giant park. In the old version of the game, the distance between you and your friend is just the number of steps you take. But in Nosa and Sosa's new version, some players have "magic shoes." If you wear a pair of shoes that shrink distances (a small "ruler"), you can reach out and grab friends who are actually far away on the map. You become a "hub," connecting to everyone easily. On the other hand, if you wear "stretchy shoes" that make distances feel huge (a large "ruler"), you might be standing right next to someone, but it feels like you are miles apart, so you don't connect.

The authors call this a "conformal deformation." In plain English, it means the shape of the space stays the same (it's still a flat sheet, a ball, or a saddle), but the scale of the space changes locally around each person. This allows the model to explain why some people are super-connected or super-isolated without having to move them to a weird spot on the map.

What They Did and Found

To test if this "stretchy ruler" idea works, the researchers ran a bunch of computer simulations and looked at eight real-world networks, including the famous friendship network of a karate club and the marriage alliances of powerful families in Renaissance Florence.

1. The Simulations: Making Messy Networks
First, they created fake networks on their computers. They started with a standard map and then added their stretchy rulers. They found that when they turned on these rulers, the fake networks looked much more like real ones. Specifically, the new model could create networks where some people had tons of friends and others had very few, even if everyone was scattered randomly on the map. The old models struggled to do this without forcing the "popular" people to cluster tightly in the center. The new model showed that you can get this "popularity" just by giving certain nodes a "shrinky" ruler.

2. The Real-World Tests: Karate Club and Florentine Families
Next, they applied their new model to real data.

The Karate Club: This is a classic dataset where a club split into two groups. The researchers found that their new model could predict who would be friends with whom better than the old models. Interestingly, they noticed that the people who were the leaders of the two factions (the "Mr. Hi" and "John A." characters) had the smallest "rulers." This meant their "magic shoes" made them feel very close to everyone else, which perfectly matched their real-life status as central, influential figures.
Florentine Families: They looked at a network of 15 families in Florence. Again, the new model did a better job of predicting the connections. It correctly identified that the Medici family (the most powerful family) had a "shrinky" ruler, making them effectively close to everyone, while other families had rulers that made them feel more distant.

3. The Results: Better Maps, But More Complexity
The researchers measured how well their new maps worked using a few different tools:

Predicting Links: The new model was better at guessing which connections existed and which didn't.
The "Vibe" of the Network: They used something called a "Laplacian spectrum" (a fancy way of measuring the overall shape and flow of the network). The new model reproduced the "vibe" of the real networks much more accurately than the old ones.
The Catch: The new model is more complex because it has to calculate a ruler for every single person. Because of this extra complexity, a standard "scorecard" for model simplicity (called the Information Criterion) sometimes preferred the old, simpler models. However, the authors argue that the new model is worth the extra complexity because it captures the messy reality of real life better.

What They Didn't Find (and What They Rule Out)

It's important to note what this paper didn't do.

It didn't prove the rulers are real: The authors are careful to say this is a statistical tool. They aren't claiming that people literally have magic shoes. They are saying that mathematically, treating people as if they have stretchy rulers helps us understand the network better.
It didn't solve community detection: When they tried to use their new model to find the two groups in the Karate club (the "factions"), it didn't actually do a better job than the old models. In fact, for this specific task, the old models sometimes worked slightly better. The new model is great at explaining why some people are popular, but it doesn't necessarily make it easier to spot the big groups.
It's not a magic bullet for everything: The authors tested this on Euclidean (flat), Spherical (ball), and Hyperbolic (saddle) spaces. While it worked well in all three, they didn't claim it works for every single type of network in the universe. They also noted that their method relies on finding the "best" ruler values through optimization, which can be tricky if the starting guess is wrong.

The Bottom Line

Nosa and Sosa have shown that by letting the "distance" between people be flexible—stretching for some and shrinking for others—we can build much better maps of complex networks. It's like realizing that in a social network, "closeness" isn't just about physical location; it's about how much effort it takes to connect. Some people make it easy (shrink the distance), and some make it hard (stretch the distance).

Their work suggests that the weird, uneven patterns we see in real life—like why some people are super-connected while others are ignored—might not be because they are in a special spot on the map, but because their personal "ruler" is different. While the math is a bit heavy, the idea is simple: in the world of connections, not all steps are created equal.

Technical Summary: Latent Space Models for Networks with Nodal Multiplicative Effects

Problem Statement
Latent space models (LSMs) represent network nodes as points in a low-dimensional geometric space, where connection probabilities are determined by the distance between these latent positions. While classical LSMs offer interpretability and parsimony, they typically assume a homogeneous global metric (Euclidean, spherical, or hyperbolic). This assumption implies that the "geometric cost" of displacement is identical across the entire latent space. Consequently, classical models struggle to capture pronounced local structural heterogeneity—such as nodes with atypical connectivity or regions of varying connection density—without altering the global geometry or introducing additive effects that modify the baseline propensity to form ties outside the distance function.

Methodology
The authors propose a generalization of classical LSMs by introducing nodal multiplicative effects motivated by a local deformation of the latent metric.

Geometric Formulation: The model approximates a conformal deformation of the metric tensor $g_\kappa$ of the ambient manifold $(M^d_\kappa, g_\kappa)$ . Instead of a constant metric, the effective distance between nodes $i$ and $j$ is modified by node-specific factors $\xi(i)$ and $\xi(j)$ .
Approximation: Recognizing that the exact geodesic distance under a conformal metric $g_\xi = \xi^2 g_\kappa$ lacks a closed-form expression, the authors derive an approximation. By fixing the geodesic of the ambient space and applying a first-order trapezoidal rule to the conformal factor along the trajectory, the effective distance is approximated as:
$\tilde{d}_\xi(z^{(i)}, z^{(j)}) \approx \frac{\xi(i) + \xi(j)}{2} d_\kappa(z^{(i)}, z^{(j)})$
Here, $\xi(i) \in \mathbb{R}^+$ acts as a local scaling factor. Values $\xi(i) > 1$ expand the effective distance (reducing connection probability), while $\xi(i) < 1$ contract it (increasing connection probability).
Statistical Model: The connection probability $p_{i,j}$ follows a logistic regression:
$p_{i,j} = \text{expit}\left( \alpha_0 - \frac{\xi(i) + \xi(j)}{2} d_\kappa(z^{(i)}, z^{(j)}) \right)$
where $\alpha_0$ is a global intercept, $z^{(i)}$ are latent positions on the manifold, and $\kappa$ determines the curvature (Euclidean, spherical, or hyperbolic).
Inference: The authors develop a hierarchical Bayesian formulation. To address the scale indeterminacy between latent positions and nodal effects, they impose the constraint $\frac{1}{n}\sum \xi(i) = 1$ $\frac{1}{n} \sum ξ (i) = 1$ . Inference is performed via a sequential optimization strategy:
1. Step 1: Estimate latent positions $Z$ and intercept $\alpha_0$ using the classical model (fixing $\xi = \mathbf{1}$ ) via Riemannian gradient ascent on the manifold.
2. Step 2: Estimate nodal effects $\xi$ by maximizing the conditional log-posterior given the fixed $Z$ and $\alpha_0$ .
  This "plug-in" approach decouples the estimation of geometry from nodal heterogeneity, improving numerical stability.

Key Contributions

Geometric Extension: The paper introduces a geometrically motivated extension to LSMs that captures node-level heterogeneity through local metric deformation rather than additive popularity effects. This preserves the global geometric structure while allowing local variations in scale.
Identifiability and Constraints: The authors establish conditions for model identifiability, addressing scale indeterminacy through a normalization constraint on $\xi$ and fixing the effective diameter of the latent space to facilitate comparison across different geometries.
Optimization Scheme: A manifold optimization scheme is developed, utilizing Riemannian gradients, exponential maps, and projections to handle the constraints of spherical and hyperbolic geometries and the simplex constraint on $\xi$ .
New Evaluation Metric: The authors propose the normalized Laplacian spectral distance as a criterion for model validation, arguing that it better evaluates the reproduction of global topological properties than edge-level prediction metrics alone.

Results
The proposed model was evaluated on synthetic data and eight real-world networks (including the Florentine families and Zachary's Karate Club).

Generative Behavior: Simulations demonstrate that nodal multiplicative effects induce greater heterogeneity in connection probabilities and degree distributions, even within homogeneous geometries like Euclidean space. The model successfully reproduces structural properties where classical models fail.
Real-World Applications:
- Florentine Families: The weighted model (with nodal effects) consistently outperformed classical models in reproducing the observed Laplacian spectrum and degree variance. The spherical model with one dimension and nodal effects achieved the smallest spectral distance.
- Karate Club: The weighted models showed superior predictive performance (higher AUC-ROC) and better structural reconstruction (lower Laplacian distance) compared to classical models. Notably, the estimated nodal effects $\xi(i)$ exhibited a strong negative correlation with standard centrality measures (degree, betweenness, closeness), indicating that nodes with high centrality effectively have contracted latent distances.
- Clustering: Interestingly, while the proposed model improved edge prediction and global topology reconstruction, it did not consistently improve community detection performance (measured by ARI, NMI, FMI) compared to classical models, suggesting the nodal effects capture node-specific propensities rather than community separation.

Significance and Claims
The authors claim that their approach increases the generative flexibility of classical latent space models. By allowing specific nodes to expand or contract their latent neighborhoods, the model captures structural heterogeneity without altering the global reference geometry. The paper asserts that a significant portion of variability traditionally attributed to latent positions can instead be captured through local geometric heterogeneity.

The work maintains the interpretability of classical LSMs while extending their expressive power. The authors position this as a step toward a broader theory of heterogeneous geometric latent representations, noting that future work could explore estimating curvature directly from data, incorporating covariates into the metric deformation, or extending the framework to weighted, directed, and temporal networks. The paper modestly notes that while the model improves structural reconstruction and prediction, it does not necessarily enhance community detection in all contexts.

The Paper's Big Idea: Stretching the Ruler

What They Did and Found

What They Didn't Find (and What They Rule Out)

The Bottom Line

More like this