Source-linked AI summary
Activity driven modeling of time varying networks
Nicola Perra, Bruno Gonçalves, Romualdo Pastor-Satorras, Alessandro Vespignani
TL;DR
Connectivity-driven network models often collapse time-varying interactions into aggregated structures that may miss instantaneous dynamics. This paper defines an empirically measurable activity potential and builds an activity-driven model, showing that heterogeneous activity produces hubs and enables analytical treatment of spreading dynamics and aggregation biases.
Problem
Time-aggregated connectivity models may not capture the instantaneous and fluctuating dynamics of time-varying networks.
Method
The paper measures nodes’ activity potential and uses its distribution to drive a dynamic-network model coupling network evolution with spreading processes.
Results
The model reproduces broad aggregated degree patterns, attributes hubs to heterogeneous activity, and derives epidemic thresholds without relying on time-aggregated connectivity.
Takeaways & Limitations
Activity distributions provide a compact basis for describing dynamic connectivity and quantifying biases in analyses of processes unfolding on evolving networks.
Takeaways & Limitations
The model does not capture link memory, relation persistence, multiple connections, or weighted links present in real network dynamics.
Abstract
from arXiv · showhide
Network modeling plays a critical role in identifying statistical regularities and structural principles common to many systems. The large majority of recent modeling approaches are connectivity driven. The structural patterns of the network are at the basis of the mechanisms ruling the network formation. Connectivity driven models necessarily provide a time-aggregated representation that may fail to describe the instantaneous and fluctuating dynamics of many networks. We address this challenge by defining the activity potential, a time invariant function characterizing the agents' interactions and constructing an activity driven model capable of encoding the instantaneous time description of the network dynamics. The model provides an explanation of structural features such as the presence of hubs, which simply originate from the heterogeneous activity of agents. Within this framework, highly dynamical networks can be described analytically, allowing a quantitative discussion of the biases induced by the time-aggregated representations in the analysis of dynamical processes.
I. RESULTS
The results define an activity potential from time-window interaction counts and show that it encodes system-level dynamics in three large-scale, time-resolved networks. An activity-driven model uses this function to generate dynamic networks, explain hubs through heterogeneous activity, and couple network dynamics with processes unfolding on them without time-scale separation.
- Activity potential: The study analyzes three large-scale, time-resolved network datasets and defines each node’s activity potential as its share of all interactions in a given time window.The activity potential is the number of interactions performed by a node divided by the total number performed by all nodes in the same window.
- Activity-driven model: The activity potential encodes system-level network dynamics and defines a process model for generating random dynamic networks.The model uses the activity potential function to characterize network structure in time.
- Activity-driven model: The model traces hubs to the heterogeneous activity of network elements and couples network dynamics with unfolding dynamical processes without time-scale separation.The results include analysis of a simple spreading process and provide explicit dynamical equations within this framework.
A. The activity potential.
The section introduces the activity potential as a time-invariant measure of agents’ interaction propensity and motivates it through time-dependent collaboration networks. It applies this characterization to scientific collaborations, Twitter messages, and film and television activity.
- A. The activity potential.: The study examines three networks with measurable individual activity: PRL author collaborations, Twitter messages, and IMDb movie and TV appearances.The corresponding activity measures are papers written, messages exchanged, and movie appearances.
- A. The activity potential.: An author’s number of collaborations depends on the observation window, as shown using PRL co-authorship networks aggregated over 1, 10, and 30 years.The comparison uses authors active during each considered time period.
- A. The activity potential.: The activity potential x_i is each agent’s interactions during a time window of length ∆t divided by all agents’ interactions in that window.The quantity is defined after measuring each agent’s activity in the three datasets.
B. Activity driven network model.
The activity-driven model generates dynamic networks from heterogeneous, time-invariant activity rates, producing instantaneous sparse structures and aggregated degree distributions that reflect activity heterogeneity. It explains hubs through high activity while omitting memory, persistence, and weighted or multiple connections.
- Model definition: Each node i receives an activity rate a_i = ηx_i, interpreted as its probability per unit time of creating new contacts.The rescaling factor η makes the average number of active nodes per unit time η⟨x⟩N.
- Model dynamics: At each time step, active vertices create m links to randomly selected vertices, after which all edges are deleted, giving interactions duration τ_i = Δt.Non-active vertices may receive links, while the network resets at the next step.
- Aggregated structure: The model produces sparse instantaneous random graphs, while aggregating connections over longer windows yields broad, skewed degree distributions.Simulations use N = 5000, m = 2, η = 10, γ = 2.8, and ε = 10^-3.
- Hub formation: Hubs arise from nodes with high activity rates that repeatedly engage in interactions, rather than from positional advantages that passively attract connections.This interpretation differs from preferential-attachment mechanisms.
- Analytical results: The integrated network’s degree distribution shares the functional form of the individual activity distribution, a relation approximately recovered in empirical data.The result is derived in the large-network, long-time limit with small k/N and k/T.
- Limitations: The model excludes link memory, social-relation persistence, multiple connections, and weighted links, which may explain discrepancies between its predictions and real networks.These effects are identified as additional features not captured by the random model.
C. Dynamical processes in activity driven networks.
The section shows that interaction timing and activity rates materially shape spreading dynamics, so epidemic thresholds should be derived from coupled network–epidemic dynamics rather than time-aggregated connectivity. Activity-driven networks provide an analytical threshold expressed on the natural interaction time scale.
- Motivation: Interaction dynamics can reverse static notions of centrality, while partner concurrency can accelerate sexually transmitted disease spread.An apparently central individual may be infected last because of interaction timing.
- Epidemic threshold: The standard reproductive number R0 = β/µ defines the epidemic threshold: epidemics can become endemic only when R0 > 1.Here β = λ⟨k⟩ is the per capita spreading rate accounting for each individual’s contact rate.
- Analytical formulation: Activity-rate analysis derives epidemic evolution equations that explicitly couple the spreading process with network dynamics.The derivation uses a general activity-potential distribution F(x) and tracks infected individuals by activity class at mean-field level.
- Analytical formulation: The mean-field update accounts for infection acquired by active susceptibles and by susceptibles contacted by infected active individuals.These mechanisms appear as separate terms in the evolution equation for each activity class.
- Epidemic threshold: The resulting epidemic threshold depends on node interaction rates, does not rely on time-aggregated connectivity, and characterizes spreading on the combined network–process time scale.The formulation incorporates each actor’s activity rate and the actual dynamics of interactions.
II. DISCUSSION
The model represents dynamical-network connectivity through an empirically measurable activity-potential distribution. It uses nodes’ activity rates to describe connectivity over time and supports analytical treatment of network processes.
- II. DISCUSSION: The model encodes the network’s connectivity pattern in a single activity-potential distribution.This function can be empirically measured in real-world networks with longitudinal data.
- II. DISCUSSION: Nodes’ activity rates define a simple dynamical process that provides a time-dependent description of network connectivity.
- II. DISCUSSION: Despite its simplicity, the model supports analytical treatment of dynamical-network processes.The supplied passage introduces this capability but ends before specifying the processes in full.
III. METHODS · A. Datasets
The study uses three empirical datasets—PRL collaborations, Twitter messages, and IMDb actor activity—to represent time-varying interactions among authors, users, and actors. Each dataset defines undirected links from observed collaborations or exchanged messages within a specified time window.
- A. Datasets: The datasets comprise collaborations in Physical Review Letters, messages exchanged on Twitter, and actor activity recorded in IMDb.These datasets support empirical analysis of author, user, and actor interaction networks.
- A. Datasets: In the PRL network, authors are nodes and coauthorship in an article creates an undirected link.Articles with more than 10 authors are excluded, and the study covers 1960–2004.
- A. Datasets: The PRL dataset records 71,583 active nodes during 1960–2004.The dataset is restricted to small collaborations whose social component is considered relevant.
- A. Datasets: The Twitter network contains users as nodes, with an undirected link when two users exchanged at least one message.The data combine over 380 million tweets from 3 million users and focus on nine months during 2008.
- A. Datasets: The Twitter time window contains 531,788 active nodes and 2,566,398 connections.The underlying user histories cover almost four years of activity, while the analyzed window spans nine months in 2008.
- A. Datasets: In the IMDb network, actors are nodes and collaboration in a movie or TV series creates an undirected link.The study covers 1950–2010 and defines activity rate as the number of movies acted in within Δt = 1 year.
- A. Datasets: The IMDb dataset records 1,273,631 active nodes and 47,884,882 connections between 1950 and 2010.The activity rate is based on each actor’s number of movies in a one-year time window.
B. Epidemic threshold
The epidemic threshold is derived by linearizing the early-stage dynamics of infectious nodes and reducing them to a closed system over the activity spectrum. The endemic state appears when the largest Jacobian eigenvalue exceeds zero, recovering the threshold in Eq. (4) under the stated parameter substitutions.
- B. Epidemic threshold: The analysis tracks the total number of infectious nodes to solve the epidemic equations.This formulation provides the starting point for the threshold calculation.
- B. Epidemic threshold: The early-epidemic approximation drops second-order terms in the activity rate a and infectious-node count It.The calculation also excludes events in which two infected nodes choose each other and uses a linear approximation in It.
- B. Epidemic threshold: Multiplying the governing equation by activity a and integrating over the activity spectrum yields an equation for θ.In the continuous-time limit, this produces a closed system of equations.
- B. Epidemic threshold: The endemic state exists when the largest Jacobian eigenvalue is greater than zero.This eigenvalue condition defines the epidemic threshold for the system.
- B. Epidemic threshold: The threshold in Eq. (4) is recovered by substituting β = λ⟨k⟩, ai = ηxi, and ⟨k⟩= 2mη⟨x⟩.These substitutions connect the derived condition to the earlier threshold expression.