Semiparametric Modeling and Analysis for Longitudinal Network Data

Yinqiu He, Jiajin Sun, Yuang Tian, Zhiliang Ying, Yang FengarXiv 2023 · DOI · arXiv

TL;DR

This paper introduces a new semiparametric framework for analyzing networks that change over time, such as bike-sharing systems or email logs. By separating static underlying relationships (latent positions) from time-varying noise (node activity levels), the researchers developed estimators that achieve "oracle" accuracy, meaning they perform as well as if the noise were already known.

What Problem Does This Paper Address?

In the world of network science, "longitudinal data" refers to interactions recorded over multiple time steps. Imagine a bike-sharing network: the physical locations of stations are static, but the activity level of each station fluctuates wildly depending on the hour of the day or the weather.

The challenge is that these two components—the static signal (how stations relate to each other geographically or socially) and the dynamic nuisance (how busy a station is at 8:00 AM vs. 11:00 PM)—are deeply entangled. Most existing models either ignore the time-varying nature of node activity or struggle to estimate the underlying network structure accurately because the number of "nuisance" parameters grows too quickly as more time points are added. The authors set out to find a way to "filter out" this time-varying noise to precisely map the network's permanent architecture.

Methods

The authors propose a Semiparametric Poisson Latent Space Model. In this model, the probability of an interaction between two nodes is determined by:

  1. Latent Vectors ($z_i$): Static coordinates in a low-dimensional space that represent the "true" relationship between nodes.
  2. Baseline Parameters ($\alpha_{it}$): Time-varying values that account for the "business" or degree heterogeneity of a node at a specific time.

To solve this, they developed two distinct estimation strategies:

  • Generalized Semiparametric One-Step Estimator: This method uses "efficient score equations" to update an initial guess. It is computationally efficient and uses quotient manifold theory to handle identifiability—a common problem in latent space models where the entire map can be rotated or reflected without changing the underlying distances.
  • Penalized Maximum Likelihood Estimator (MLE): This approach uses a convex surrogate (nuclear norm regularization) to handle the low-rank nature of the network without needing to pre-specify the exact number of dimensions in the latent space.

Datasets

The primary real-world application uses the New York Citi Bike Dataset from August 2019.

  • Scale: Over 2.3 million rides across 782 stations over a 24-hour period.
  • Context: Each station is a node, and the number of rides between stations in a given hour represents the edge weight.
  • Benchmark: The authors compared their results against the COSIE (Common Subspace Independent Edge) model, showing that their semiparametric approach better captured the actual geographic relationships of the city.

Key Insights and Ideas

  • The "Inverse Proportional to T" Property: One of the most significant findings is that the estimation error for the static latent positions decreases at a rate of $1/T$ (where $T$ is the number of time points). This means that as you collect more data over time, your map of the network's "permanent" structure becomes significantly more accurate, even if the new data is very noisy.
  • Manifold Geometry: The authors treat the network not as a flat Euclidean space, but as a quotient manifold. This allows the math to ignore "redundant" information (like rotations of the latent map) and focus purely on the intrinsic geometry of node relationships.
  • Nuisance doesn't have to be a Nuisance: By using semiparametric efficiency theory, the authors proved that you can achieve the "oracle" error rate—the best possible theoretical accuracy—even when you don't know the baseline activity levels of the nodes.

Key Results

  • Geographic Recovery: When applied to the NYC Citi Bike data, the model's 2D latent positions closely mirrored the actual map of New York. It clearly separated Manhattan stations from those in Brooklyn and Queens, identifying the East River as a natural barrier that reduces interactions.
  • Oracle Performance: In simulations, both the one-step estimator and the penalized MLE achieved the predicted oracle convergence rates. Even when time-varying heterogeneity was high, the model successfully extracted the static latent signal.
  • Robustness: The model proved effective even when the number of time points ($T$) was much larger than the number of nodes ($n$), a scenario that typically causes other models to fail due to parameter over-fitting.

Limitations and Future Work

The authors acknowledge several areas for expansion:

  • Directed Networks: The current model assumes interactions are undirected (A to B is the same as B to A). Future iterations could incorporate directional data (e.g., morning commutes vs. evening returns).
  • Sparsity: The model currently treats the network as relatively dense. Adapting the framework for very sparse networks (where most nodes never interact) remains an open challenge.
  • Dynamic Latent Spaces: While this paper focuses on static latent positions, some networks have structures that evolve slowly over months or years. The authors suggest that combining their baseline activity adjustment with a slowly shifting latent space could be a powerful next step.

Summary updated

Abstract

We introduce a semiparametric latent space model for analyzing longitudinal network data. The model consists of a static latent space component and a time-varying node-specific baseline component. We develop a semiparametric efficient score equation for the latent space parameter by adjusting for the baseline nuisance component. Estimation is accomplished through a one-step update estimator and an appropriately penalized maximum likelihood estimator. We derive oracle error bounds for the two estimators and address identifiability concerns from a quotient manifold perspective. Our approach is demonstrated using the New York Citi Bike Dataset.
BibTeX
@misc{he2023semiparametric,
  author = {Yinqiu He and Jiajin Sun and Yuang Tian and Zhiliang Ying and Yang Feng},
  title = {{Semiparametric Modeling and Analysis for Longitudinal Network Data}},
  year = {2023},
  eprint = {2308.12227},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.12227},
}
Figure