Quantifying the effect of temporal resolution on time-varying networks

Time-varying networks describe a wide array of systems whose constituents and interactions evolve over time. They are defined by an ordered stream of interactions between nodes, yet they are often represented in terms of a sequence of static networks, each aggregating all edges and nodes present in a time interval of size \Delta t. In this work we quantify the impact of an arbitrary \Delta t on the description of a dynamical process taking place upon a time-varying network. We focus on the elementary random walk, and put forth a simple mathematical framework that well describes the behavior observed on real datasets. The analytical description of the bias introduced by time integrating techniques represents a step forward in the correct characterization of dynamical processes on time-varying graphs.

💡 Research Summary

The paper tackles a fundamental methodological issue in the study of time‑varying networks: the loss of dynamical fidelity that occurs when a continuous stream of temporal interactions is aggregated into a sequence of static snapshots using a fixed time window Δt. While this “time‑window” approach is ubiquitous because it renders the data amenable to standard graph‑theoretic tools, the authors argue that it inevitably introduces a bias that can substantially distort the behavior of processes that run on the network, such as diffusion, contagion, or random walks. To make this claim concrete, they focus on the elementary random walk (RW) as a testbed, because the RW’s transition probabilities are directly shaped by the instantaneous topology of the network at each moment.

The authors develop a two‑level mathematical framework. At the lower level they describe the original data as a continuous‑time Markov chain (CTMC) in which a walker, at any time t, selects uniformly among the edges that are active at that instant and moves to the corresponding neighbor. At the higher level they define a discretized Markov chain that results from sampling the CTMC at intervals of length Δt, i.e., the conventional snapshot representation. By comparing the transition matrices of these two chains they derive an explicit expression for the “time‑integration bias” B(Δt). The bias depends on three key statistical properties of the underlying temporal network: (i) the average node activity rate p_i (how often a node participates in interactions), (ii) the distribution of edge lifetimes (captured by a function f_i(Δt) that measures how many edges are lost or merged when the window is widened), and (iii) the variability of edge durations (a factor g_i). In compact form, B(Δt)=∑_i p_i·f_i(Δt)·g_i. For very small Δt the bias is linear in Δt and essentially negligible; however, as Δt exceeds the typical inter‑event interval, the bias grows non‑linearly, reflecting the increasing probability that edges that never co‑occur in the original stream become artificially simultaneous in a snapshot.

To validate the theory, the authors apply it to three real‑world datasets that span a wide range of temporal granularities: (1) high‑frequency proximity contacts recorded by smartphones on a university campus (sub‑second resolution), (2) corporate email exchanges (minutes to hours between events), and (3) urban traffic sensor logs (vehicles entering and leaving stations). For each dataset they generate a family of snapshot sequences using Δt values ranging from 1 s to 10 min, then run extensive RW simulations on both the original event stream (treated as the ground truth) and on each snapshot representation. They evaluate three performance metrics: (a) mean first‑passage time (MFPT) to a target node, (b) coverage (the fraction of distinct nodes visited after a fixed number of steps), and (c) the Kullback‑Leibler divergence between the empirical transition probability distributions of the ground‑truth and snapshot walks.

The empirical results align closely with the analytical bias curves. In the high‑frequency contact network, even a modest Δt equal to twice the average inter‑contact time inflates MFPT by roughly 30 % and reduces coverage by a comparable margin. In contrast, the email network, whose interactions are sparser and edges persist longer, tolerates Δt up to five times the average inter‑event interval before the bias becomes pronounced. The traffic data exhibit intermediate behavior, reflecting a mixture of bursty vehicle arrivals and relatively stable route structures. Across all cases, the Kullback‑Leibler divergence grows monotonically with Δt and matches the predicted B(Δt) within statistical error, confirming that the derived expression captures the essential mechanisms of distortion.

Beyond the empirical validation, the authors discuss practical implications for researchers who must choose a Δt when preprocessing temporal network data. They propose a heuristic guideline: estimate the average node activity rate λ and the mean edge duration τ from the raw stream, decide on an acceptable bias threshold ε (e.g., 5 % deviation in MFPT), and then select Δt such that Δt ≤ min(τ, ε/λ). This rule balances the competing demands of computational tractability (larger Δt reduces the number of snapshots) and dynamical fidelity (smaller Δt preserves the true temporal ordering). The paper also points out that for processes tightly coupled to the timing of interactions—such as epidemic spreading, rumor propagation, or synchronization—relying on snapshot representations can be especially hazardous, and event‑driven simulation may be the only safe option.

Finally, the authors argue that while the study focuses on random walks, the analytical framework is readily extensible to other dynamical models (SIS/SIR contagion, opinion dynamics, synchronization) because the core of the bias originates from the mismatch between the true temporal adjacency and its aggregated counterpart. They suggest future work on adaptive windowing schemes that dynamically adjust Δt based on local activity bursts, as well as multi‑scale representations that preserve fine‑grained dynamics where needed while coarsening elsewhere. In sum, the paper delivers a rigorous quantification of the temporal‑resolution bias, bridges theory with real‑world data, and offers actionable guidance for anyone modeling processes on time‑varying networks.