Hubs, Rich Club, Scale Free Network IV124 Josef Spurný & Eva Výtvarová Faculty of Informatics, Masaryk University March 31, 2023 Introduction Hubs Short definition: nodes with high degree What does high mean: reminder: binomial degree distribution in random network hubs: far to the right from the expected distribution far means, for example, at least one standard deviation from the mean J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 2 / 32 Introduction Hubs: A ZOO of Complex Networks 1 1 https://noduslabs.com/radar/ types-networks-random-small-world-scale-free/ J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 3 / 32 Introduction Hubs: Motivation Why do we observe hubs? network structure is a result of self-organization real-world systems have limited resources maintaining links is costly hubs allow coordination, faster spreading... hubs are found across multiple scales - fractal (scale-free) distribution scale-free distribution is a result of optimization process in systems with finite resources2 2 Csermely, P. (2006), pp. 20. Weak links. J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 4 / 32 Introduction Hubs: Motivation model (network) clusters small diameter hubs Grid yes no no Erdös-Rényi (random) no yes no Watts-Strogatz (small world) yes yes no Barabási-Albert (scale-free) no yes yes Many real-world networks contain hubs: protein-protein interaction, gene expression, metabolic networks human communication (phone calls, emails...) human interaction (science / movie cooperation, wealth distribution...) www, internet, power grids J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 5 / 32 Introduction Guimera et al. (2007) doi:10.1038/nphys489 J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 6 / 32 Introduction Hubs in Detail If the network has a community structure, we can distinguish between: global hubs provincial hubs within modules connector hubs connecting multiple modules peripheral hubs satellite connectors We quantify this using the so-called participation coefficient. J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 7 / 32 Introduction Participation Coefficient3 Pi = 1 − NM s=1 κis ki 2 NM is the total number of modules κis is the number of edges from node i to module s P ≤ 0.3 provincial hubs 0.3 < P ≤ 0.75 connectors 0.75 < P global hubs 3 Guimera, Amaral (2008). Functional cartography of complex metabolic networks. J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 8 / 32 Rich-club Rich-club (a) (b) J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 9 / 32 Rich-club Rich-club (k-core) Rich club is a result of network assortativeness (homophily) Describes whether dominant nodes form a tightly interconnected core. φ(k) = 2E>k N>k(N>k − 1) E>k is the number of edges between N>k nodes with degree greater than k it represents the fraction of edges between these nodes out of all possible edges J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 10 / 32 Rich-club Rich-club: Normalization Null model: φun(k) ∼ k2 ⟨k⟩N Normalized RC: ρunc(k) = φ(k) φun(k) J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 11 / 32 Rich-club Rich-club examples4 4 Colizza et al. (2006) doi:10.1038/nphys209 J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 12 / 32 Rich-club Detour Project proposals... J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 13 / 32 Power law Hubs are nodes with a surprisingly high degree. What does the surprisingly high degree mean? J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 14 / 32 Power law Hubs and node degree distribution random network vs. scale-free network binomial distribution Gaussian distribution Poisson distribution normal distribution power-law distribution Pareto distribution heavy-tailed distribution fat-tailed distribution J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 15 / 32 Power law Logarithmic representation of distribution On a normal histogram, we don’t see much, but a log-log representation is much more interesting. J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 16 / 32 Power law Power law distribution5 5 Newman, M. E. (2005) DOI:10.1080/00107510500052444 J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 17 / 32 Power law Power law distribution J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 18 / 32 Power law Power law distribution J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 19 / 32 Power law From graph to equation: Power law log(ak) = k · log(a) log(ab) = log(a) + log(b) The relationship can be described as y = kx + q: log(P) = log(c) − γ log(D) After exponentiation : P = cD−γ where P is the probability to meet a good friend or acquaintance c is constant γ is scaling exponent J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 20 / 32 Power law Power law: variance The second moment (variance) is generally infinite =⇒ as the sample (network) size increases, so will the maximum value. k σ yeast protein-protein interaction network 2.9 4.88 E. Coli metabolic network 5.58 20.79 WWW network 4.60 30.27 J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 21 / 32 Power law Let’s compute the scaling exponent: Log-log representation6 Problem: as the degree increases, fewer samples are available, hence noise increases. 6 Note: always use logarithm with base 10. J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 22 / 32 Power law Let’s compute the scaling exponent: Log-log representation Solution: logarithmic bins J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 23 / 32 Power law Let’s compute the scaling exponent: Log-log representation Solution: complementary cumulative distribution function P≥(x) = x1−γ J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 24 / 32 Scale-free networks Scale-free networks A scale-free network is a network whose degree distribution follows a power law, at least asymptotically. Many systems seem to be in the regime of 2 < γ < 3. Is there any reason for that? J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 25 / 32 Scale-free networks Example: Scale-free Search Strategy When social insects seek food, they frequently make cost-efficient small trips From time to time, they make longer jumps Rarely, they make very long journeys This search strategy is called Lévy flight and follows the power law P = cL−α, where L is length of a trip Why it is the best strategy? It minimizes probability to return to the same site (disadvantage of random search) It maximizes chance to end in new location (disadvantage of grid search) =⇒ best survival strategy J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 26 / 32 Scale-free networks Lévy flight Lévy flight search patterns for 1000 steps. Values for α = 2 proven to be optimal7. 7 Csermely (2006). Weak links, pp. 29 J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 27 / 32 Scale-free networks Classes of scale-free networks Anomalous regime γ ≤ 2 the degree of the largest hub grows faster than the size of the network such a network is not asymptotically possible without loops Scale-free regime 2 < γ < 3 the first moment of the distribution (mean) is finite, other moments diverge (variance, skewness, ...) specific behavior of dynamic processes (diffusion) robustness against random failure vulnerable to targeted attack (unlike a random network) Random network regime γ > 3 probability of large hubs decreases too fast hard to distinguish from a random network J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 28 / 32 Scale-free networks Classes of scale-free networks J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 29 / 32 Scale-free networks Empirical estimation of exponent Motivation: important for null models, e.g. for the study of dynamic processes Procedure: start with log-log transformation, fit a line use logarithmic intervals or cumulative distribution to remove noise apply the method of least squares, maximum likelihood, etc. in your favorite statistical software J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 30 / 32 Scale-free networks Example in Excel ... J. Spurný, E. Výtvarová ·Hubs ·March 31, 2023 31 / 32