Statistical inference of dynamical processes on networks

Technical Summary: Statistical Inference of Dynamical Processes on Networks

Problem Statement
Understanding the mechanisms driving spreading processes on networks—ranging from disease transmission and opinion dynamics to information diffusion and ecological cascades—is central to network science. While empirical data frequently reveals the network structure (who interacts with whom), it rarely provides direct information regarding the functional rules governing how interactions drive state changes. Consequently, the underlying interaction mechanisms must be inferred from observed dynamics by selecting among competing hypotheses. This paper addresses the inverse problem of statistically inferring the interaction rules (transition rates) of binary-state spreading processes on networks, given known structure and observed node state trajectories. A critical challenge is determining the conditions under which distinct mechanistic models are statistically distinguishable, particularly when different models can produce similar macroscopic behaviors.

Methodology
The authors propose a general framework for model selection based on the likelihood ratio between a true generative model and a set of competing hypotheses. The study focuses on binary-state dynamics (nodes in state 0 or 1) where transition probabilities depend on the node's degree ($k$) and the number of infected neighbors ($m$).

1. Theoretical Framework:

   • The authors define a family of six archetypal spreading mechanisms: Independent, Simple, Voter, Ising, Threshold, and Majority. These are parameterized by infection ($F$) and recovery ($R$) rates.
   • They derive the expected log-likelihood ratio between two models, showing it is a function of the Kullback-Leibler (KL) divergence between the transition probabilities of the models, weighted by the observed "dynamical range" (the distribution of time-aggregated transitions across $(k, m)$ classes).
   • A key theoretical contribution is the definition of asymptotic detectability ($\lambda^*$). By utilizing Approximate Master Equations (AMEs)—which describe the evolution of node state fractions in the thermodynamic limit ($N \to \infty$)—the authors derive an analytical approximation of detectability. They demonstrate that this asymptotic measure accurately predicts inference outcomes in finite systems, stabilizing as system size increases.

2. Systematic Exploration:

   • The authors systematically explore a high-dimensional parameter space ($\Omega$) encompassing model parameters, network topology (average degree $\langle k \rangle$, degree heterogeneity), and initial conditions.
   • They distinguish between transient and stationary dynamics, noting that the detectability of a mechanism depends heavily on whether the observation period captures the system's relaxation or its steady state.

3. Empirical Application:

   • The framework is applied to a diverse set of seven empirical datasets, including game-theoretic experiments, social media activity (Higgs boson discovery), urban traffic flow, bike-sharing systems, supply chain dynamics, and primate behavioral interactions.
   • A crucial aspect of the empirical analysis is the investigation of data preprocessing. The authors vary temporal aggregation windows ($\Delta t$) and feature binarization thresholds ($\phi$) to assess how coarse-graining affects model selection. They also introduce "noisy" model variants, where a fraction of transitions is modeled as independent (random) noise, to account for unobserved factors in real-world data.

Key Results

• Detectability and System Size: Asymptotic detectability ($\lambda^*$) stabilizes with increasing system size, validating its use as a predictor for finite systems. However, model selection can fail even in large systems if the dynamical range does not sufficiently explore the parameter space where mechanisms differ.
• Network Density and Sparsity: Detectability generally increases in sparse networks (low average degree). As network density increases, the parameter regions where mechanisms are distinguishable shrink, and different mechanisms tend to produce similar dynamical ranges, leading to higher confusion between models.
• Phase Transitions: Inference is most successful near phase transitions (e.g., the epidemic threshold in SIS dynamics). At these critical points, the system exhibits critical slowing down, generating a larger volume of observable transitions and a broader dynamical range, which enhances the ability to distinguish between models.
• Impact of Preprocessing: The choice of temporal resolution and binarization thresholds fundamentally reshapes model selection outcomes. A single dataset can favor different mechanisms (e.g., simple vs. threshold contagion) depending on the aggregation scale.
• Empirical Findings: In most empirical datasets, the Independent model (random transitions) or Noisy Simple models (a mix of independent noise and simple contagion) provide the best fit. Pure mechanistic models (like Threshold or Ising) are rarely the sole best fit without accounting for noise. The "noisy" variants consistently outperform single-mechanism models, suggesting that real-world dynamics often involve a combination of structural influence and exogenous/random factors.

Significance and Claims
The paper claims to be the first systematic effort to assess both the limits of detectability and the prevalence of paradigmatic mechanisms in networked spreading phenomena.

• Theoretical Insight: The work demonstrates that focusing on the "dynamical range" (how data is distributed over structural classes) is more informative than simply increasing data volume. It reveals that conceptually different models can have similar statistical behavior depending on the dynamical regime, creating a "landscape of mechanisms" where multiple models may have comparable explanatory power.
• Methodological Advance: By showing that asymptotic approximations (AMEs) can reliably inform model selection in finite systems, the authors provide a tool to identify parameter regions where inference is likely to fail before conducting expensive simulations or experiments.
• Practical Implication: The results highlight that data preprocessing is not merely a technical step but a modeling assumption that can reframe the underlying process. The prevalence of "noisy" models in empirical data suggests that simple, idealized mechanisms are often insufficient to describe real-world phenomena without accounting for unobserved heterogeneity or exogenous drivers.

The authors conclude that statistical model selection can fail under common conditions (dense networks, specific parameter regimes, or inappropriate data aggregation) and suggest that future interpretations of empirical spreading phenomena must carefully consider the interplay between network structure, observation scales, and the inherent limitations of distinguishability between competing mechanisms.