Memory in network flows and its effects on spreading dynamics and community detection

Random walks on networks is the standard tool for modelling spreading processes in social and biological systems. This first-order Markov approach is used in conventional community detection, ranking, and spreading analysis although it ignores a potentially important feature of the dynamics: where flow moves to may depend on where it comes from. Here we analyse pathways from different systems, and while we only observe marginal consequences for disease spreading, we show that ignoring the effects of second-order Markov dynamics has important consequences for community detection, ranking, and information spreading. For example, capturing dynamics with a second-order Markov model allows us to reveal actual travel patterns in air traffic and to uncover multidisciplinary journals in scientific communication. These findings were achieved only by using more available data and making no additional assumptions, and therefore suggest that accounting for higher-order memory in network flows can help us better understand how real systems are organized and function.

💡 Research Summary

The paper challenges the prevailing assumption in network science that random walks can be adequately modeled as first‑order Markov processes (M1), where the next step depends only on the current node. By constructing second‑order Markov models (M2) that incorporate the immediately preceding node, the authors demonstrate that many real‑world flow systems exhibit strong memory effects that are invisible to M1 models. Using pathway data from six diverse domains—U.S. airline itineraries, aggregated city traffic, journal citation chains, patient movements between hospital wards, GPS‑tracked taxis, and email forwarding—they build “memory networks” where each memory node represents a directed link (i→j) and transitions correspond to linked pairs (i→j → j→k). Transition probabilities are estimated directly from observed conditional flows.

Entropy‑rate analysis shows that M2 models reduce conditional entropy by 1–2 bits relative to M1, implying a substantial decrease in effective connectivity (equivalent to over‑estimating the number of neighbors by 200‑400 % in an unweighted network). Nodes with the strongest memory effect, such as the Las Vegas hub airport, display dramatically higher two‑step return rates (up to eightfold) and a pronounced shift from high‑entropy M1 behavior to low‑entropy M2 behavior.

When the Infomap community‑detection algorithm is applied to the M2 transition matrix, the resulting modules are more granular and align better with known geographic or topical structures. For example, the U.S. air‑traffic network splits into distinct Las Vegas and Atlanta modules, reflecting passengers’ tendency to return to their origin. In the citation network, multidisciplinary journals emerge as separate communities, a pattern missed by M1‑based clustering.

Ranking using PageRank on M2 also yields scores that correlate more closely with real influence: airports and journals that serve as frequent “memory nodes” (i.e., appear often as the second step in observed paths) receive higher, more realistic ranks than in the M1 case.

Epidemic spreading simulations (SIR) on M1 and M2 versions of the networks reveal only minor differences in total infection size, suggesting that disease dynamics are dominated by long‑range connections where two‑step memory has limited impact. In contrast, information diffusion processes (e.g., retweets, email forwards) are markedly sensitive to memory; M2 models produce faster, broader spread, highlighting the contextual nature of human communication.

Statistical significance is established through bootstrap resampling of pathways and surrogate data tests; most nodes and networks (except patient and email data, which suffer from limited sample size) show a significant second‑order effect that cannot be attributed to noise.

The authors conclude that incorporating memory into network flow models is essential for accurate community detection, ranking, and information‑spreading analysis. Since the required data are often already available, constructing M2 models incurs little extra cost while delivering substantial insight. They suggest future work on higher‑order (third‑order and beyond) models, real‑time estimation of memory networks, and leveraging memory effects for optimization in routing, recommendation, and control of spreading processes.