Source-linked AI summary
Persistence and periodicity in a dynamic proximity network
Aaron Clauset, Nathan Eagle
TL;DR
Dynamic social-network data are often reduced to fixed-window snapshots, but the choice of window can bias measured structure. Using highly resolved proximity data from 66 individuals, the paper characterizes persistence and periodicity, shows that snapshot-based statistics depend strongly on the rate, and proposes a natural rate for discretization.
Problem
Fixed-window snapshots can bias network-structure estimates because the chosen snapshot rate determines how continuous temporal variation is represented.
Method
The study analyzes continuously varying proximity edges, their persistence, temporal statistics, and snapshot sequences constructed at multiple rates.
Results
The topology evolves across broad timescales with strong calendar-driven periodicities, and the natural snapshot rate is Δ_nat = 4.08 hours.
Takeaways & Limitations
Dynamic proximity networks are multi-scale, and discretization should use a natural rate that smooths high-frequency variation while preserving lower-frequency structure.
Takeaways & Limitations
The natural rate may vary substantially across social contexts; more physically active environments may have a faster natural timescale than this office-based network.
Abstract
from arXiv · showhide
The topology of social networks can be understood as being inherently dynamic, with edges having a distinct position in time. Most characterizations of dynamic networks discretize time by converting temporal information into a sequence of network "snapshots" for further analysis. Here we study a highly resolved data set of a dynamic proximity network of 66 individuals. We show that the topology of this network evolves over a very broad distribution of time scales, that its behavior is characterized by strong periodicities driven by external calendar cycles, and that the conversion of inherently continuous-time data into a sequence of snapshots can produce highly biased estimates of network structure. We suggest that dynamic social networks exhibit a natural time scale Δ_{nat}, and that the best conversion of such dynamic data to a discrete sequence of networks is done at this natural rate.
Network Statistics
The paper converts continuous proximity intervals into snapshots using a chosen rate Δ, then measures degree, clustering, and adjacency overlap. These measures are affected by temporal resolution, with larger windows averaging out higher-frequency fluctuations.
- Snapshot construction: A snapshot rate Δ partitions proximity intervals into network snapshots covering successive time windows.An edge is included when two nodes are proximate at any time during a window.
- Snapshot construction: Small Δ values produce sparse snapshots containing only a few edges.This sparsity distinguishes the snapshots from many conventional social-network representations.
- Network measures: The analysis measures mean degree, clustering coefficient, and adjacency correlation across the snapshot sequence.Adjacency correlation captures the overlap between a vertex’s neighbor sets in two snapshots.
- Temporal resolution: Larger snapshot rates average out higher-frequency fluctuations in mean degree, clustering, and adjacency correlation.The comparison uses snapshot rates from 5 to 1440 minutes.
Network Analysis
The study examines how snapshot rate changes observed network dynamics in highly resolved proximity data. It finds broad persistence timescales, strong periodicities, and network statistics that vary systematically with the chosen rate, motivating a natural rate of 4.08 hours.
- Persistence and timescales: 22.83 minutes is the average edge persistence, while three edges persist longer than 1440 minutes.These durations span fast and slow topological variation across a broad range of timescales.
- Snapshot-rate effects: The 5-minute snapshots retain high-frequency noise overlaid on structure associated with the daily work cycle.Increasing Δ progressively averages out these high-frequency variations.
- Snapshot-rate effects: The measured network statistics grow essentially monotonically with snapshot rate Δ, so the choice of Δ determines their values.Mean degree and clustering increase with Δ, whereas adjacency correlation is expected to decrease.
- Natural rate: Strong daily spectral peaks support a natural snapshot rate of Δ_nat = 4.08 hours.At this rate, the reported natural statistics are ⟨k⟩nat = 2.24, Cnat = 0.084, and γnat = 0.88.
Discussion
Snapshotting continuous-time social networks can bias measured structure because the window length determines which temporal variation is retained. The study therefore proposes a natural rate that smooths high-frequency variation while preserving lower-frequency patterns, while noting that this rate depends on social context.
- Incorrectly choosing the snapshot rate can impose strong bias on analyses and conclusions about dynamic social systems.
- The three metric autocorrelations fall to zero after 6.08, 5.25, and 6.25 hours, respectively.
- 24- and 12-hour peaks dominate the metric power spectra, with additional 8-hour peaks for mean degree and clustering coefficient.
- The proposed natural snapshot rate smooths high-frequency variation while preserving important low-frequency structural patterns.
- The calculated natural rate is Δ_nat = 4.08 hours and is presumably related to the day and human work day.
- The network topology evolves across broad time scales, with variation likely driven by external calendar periodicities such as day, week, month, and season.