TL;DR
This paper introduces a novel framework to simultaneously analyze the temporal and topological properties of evolving networks, revealing distinct patterns between virtual and physical interactions. The findings suggest that temporal-topological correlations are stronger in virtual networks, potentially due to lower interaction costs facilitating social contagion.
What Problem Does This Paper Address?
Real-world systems, from social interactions to communication networks, are dynamic and can be represented as evolving networks where connections change over time. Understanding these networks requires analyzing both when interactions occur (temporal properties) and how they are structured (topological properties). While previous work has often focused on one aspect or the other, a systematic method to characterize the interplay between temporal and topological relationships within these evolving networks has been missing. This paper aims to fill that gap by developing and applying a method to quantify how closely timed events are also topologically close, and vice versa.
Methods
The authors propose a three-pronged approach to characterize temporal-topological properties:
- Global Activity Correlation: They analyze the "activity time series" of the entire network, which records the total number of contacts at each time step. By aggregating this series over different time bins, they measure how correlated the network's activity is over time.
- Interplay of Topological and Temporal Distances: They investigate the relationship between the time difference and the topological distance (shortest path) between any two contacts. This helps understand if contacts that happen close in time are also between nodes that are close in the network structure.
- Local Neighborhood Correlation: They examine "ego-networks" (local neighborhoods centered around a link) to see if contacts within these local structures are temporally correlated, forming "event trains."
To ensure their findings are robust, they also introduce three control randomization methods that systematically preserve or remove specific temporal and topological correlations, allowing for comparison against null models.
Datasets
The study utilizes 13 real-world temporal networks, categorized into two main types:
- Virtual Contact Networks:
- DNC Mail Part 2
- Manufacturing Email (ME)
- College Messages (CM)
- Email EU (EEU)
- Physical Contact Networks:
- Infectious (Science Gallery, Dublin)
- Primary School (PS)
- High School 2012 (HS2012)
- High School 2013 (HS2013)
- Hypertext 2009 (HT2009)
- SFHH Conference
- Workplace (WP)
- Workplace (WP2)
- Hospital
For each dataset, they focus on the largest connected component and remove long periods of inactivity.
Key Insights and Ideas
- Virtual vs. Physical Divide: A striking difference emerges between virtual and physical contact networks. Temporal-topological correlations are consistently stronger in virtual networks.
- Cost of Interaction Matters: The authors hypothesize that the lower cost and easier accessibility of virtual interactions (like emails or online messages) facilitate "social contagion" – the spread of activity or influence between connected individuals.
- Local Events Form Trains: Contacts within local neighborhoods (ego-networks) tend to form "event trains" (bursts of activity), indicating temporal clustering.
- Spatial Constraints in Physical Networks: In physical contact networks, the number of distinct links involved in an event train seems to reflect spatial constraints. For example, school and workplace networks show different patterns, likely due to differing levels of spatial freedom.
- Ego-network Trains are Not Just Single-Link Activity: Event trains in ego-networks are generally composed of activities from multiple links, not just a single one, suggesting coordinated local activity.
Key Results
- Temporal-Topological Correlation: The study consistently shows that contacts close in time are also close in topology across all analyzed networks. This trend is significantly more pronounced in virtual networks compared to physical ones.
- Event Trains: The train size distribution analysis reveals that real-world networks exhibit longer event trains than randomized networks, indicating significant temporal correlation in local events.
- Activity Balance: An "activity balance" measure shows that events within an ego-network train are typically contributed by both connected nodes, but this balance is less pronounced (more unbalanced) in virtual networks, suggesting a tendency for individuals to contact multiple others simultaneously in virtual settings.
Limitations and Future Work
- Undirected Networks: The current methods are confined to undirected networks. The authors suggest that developing characterization methods for directed temporal networks would be a valuable future direction.
- Dynamic Processes: The paper highlights that understanding detected temporal-topological properties could enhance the study of dynamic processes (like information diffusion) on these networks.
- Realistic Network Models: The identified patterns and differences are expected to inspire the development of more realistic temporal network models that can jointly reproduce temporal and topological contact properties.
Summary updated