Abstract
Models of spread processes over non-trivial networks are commonly motivated by modeling and analysis of biological networks, computer networks, and human contact networks. However, identification of such models has not yet been explored in detail, and the models have not been validated by real data. In this paper, we present several different spread models from the literature and explore their relationships to each other; for one of these processes, we present a sufficient condition for asymptotic stability of the healthy equilibrium, show that the condition is necessary and sufficient for uniqueness of the healthy equilibrium, and present necessary and sufficient conditions for learning the spread parameters. Finally, we employ two real datasets, one from John Snow's seminal work on cholera epidemics in London in the 1850's and the other one from the United States Department of Agriculture, to validate an approximation of a well-studied network-dependent susceptible-infected-susceptible (SIS) model.
Introduction
While researchers have developed numerous mathematical models for how viruses, information, or behaviors spread across networks, a significant gap remains: these models are rarely validated with real-world time-series data. Most existing work focuses on either purely theoretical stability analysis or uses real data only to build the network structure while simulating the actual spread process with arbitrary parameters. This research addresses the challenge of system identification: given data on a spread process and a known network structure, can we uniquely determine the infection and healing rates? Furthermore, do these theoretical models actually reflect reality?
Modeling Framework and Relationships
The core of this work focuses on a discrete-time Susceptible-Infected-Susceptible (SIS) model, which serves as a mean-field approximation of a $2^n$-state Markov chain. A key theoretical contribution is the clarification of how various popular models in the literature relate to one another.
As illustrated in the hierarchy above, the "True System" can be modeled as a $2^n$ state model, which simplifies to a continuous-time mean-field approximation. By applying Euler’s Method or first-order truncation to different discrete-time formulations, the authors demonstrate that several popular discrete-time SIS models are mathematically equivalent under specific conditions (e.g., when the sampling interval $h=1$).
Stability and Parameter Identification
The authors provide a rigorous mathematical framework for analyzing these models:
- Stability Analysis: Using Lyapunov stability theory, they establish that the "healthy state" (where the infection is eradicated) is the unique equilibrium if the spectral radius of the system matrix is less than or equal to one.
- Parameter Identification: They derive necessary and sufficient conditions to learn the infection rate ($\beta$) and healing rate ($\delta$) from observed data. They show that with as few as two time steps of full-state measurements, these parameters can be recovered using a pseudo-inverse approach, provided the system has sufficient "diversity" in its nodal states. Notably, even if the exact sampling interval ($h$) is unknown, the ratio of infection to healing ($\beta/\delta$)—which determines long-term behavior—can still be recovered exactly.
Empirical Validation
The paper employs two distinct datasets to validate the SIS model approximation:
John Snow’s Cholera Map (1854)
The authors digitized Snow's famous map of the Soho district. Because the original data was a spatial snapshot, they used statistical methods and assumptions about household sizes to reproduce the time-series data of deaths. Interestingly, a standard "nearest neighbor" network did not fit the data well. However, when the network was modified to show the Broad Street pump affecting all households (reflecting the true source of the outbreak), the model's output closely matched historical records, capturing the general shape and peak of the epidemic.
USDA ACRE Program (2008-2013)
This dataset tracks payouts for the Average Crop Revenue Election program, treating the adoption of this complex agricultural subsidy as a "contagion" spreading through a network of U.S. counties. The authors learned spread parameters using data from Idaho and then simulated adoption across the entire contiguous United States. The simulation achieved a scaled Frobenius norm error of 0.2348. As predicted by the irreversible nature of the program, the learned healing rate ($\delta$) was significantly smaller than the infection rate ($\beta$), reflecting that once farmers adopted the subsidy, they rarely reverted.
Discussion and Conclusion
This study demonstrates that contagion is a universal concept; the same mathematical tools used for biological viruses can effectively model social contagions, such as farmers adopting new financial technologies.
However, limitations remain. The current parameter identification technique assumes noise-free measurements, and future work must account for the inherent noise in real-world data collection. Additionally, while this study assumes the underlying network structure is known, many real-world scenarios require the network itself to be inferred alongside the spread parameters. Future research will also need to explore the stability and identification of "endemic" states where infections persist indefinitely.
Summary updated