Abstract
Here we report a comprehensive analysis of the robustness of seven highquality realworld complex weighted networks to errors and attacks toward nodes and links. We use measures of the network damage conceived for a binary (e.g. largest connected cluster LCC, and binary efficiency Eff_{bin}) or a weighted network structure (e.g. the efficiency Eff, and the total flow TF). We find that removing a very small fraction of nodes and links with respectively higher strength and weight triggers an abrupt collapse of the weighted functioning measures while measures that evaluate the binarytopological connectedness are almost unaffected. These findings unveil a problematic responsestate where the attack toward a small fraction of nodeslinks returns the realworld complex networks in a connected but inefficient state. Our findings unveil how the robustness may be overestimated when focusing on the connectedness of the components only. Last, to understand how the networks robustness is affected by link weights heterogeneity, we randomly assign link weights over the topological structure of the realworld networks and we find that highly heterogeneous networks show a faster efficiency decrease under nodeslinks removal: i.e. the robustness of the realworld complex networks against nodeslinks removal is negatively correlated with link weights heterogeneity.
Introduction
The robustness of a network is its ability to maintain the system functioning in case of failures of nodes or links. Networks robustness is extremely important and has been widely investigated in last years in different fields of science^{1,2,3,4,5,6,7,8,9,10,11,12,13}. A comprehensive analysis of network robustness considers the failure of both nodes (e.g. Iyer et al.^{9}) and links (e.g. Pajevic and Plenz^{13}). Initially, large attention has been dedicated to binarytopological analyses where the links among nodes are only present or absent. Yet, recent studies evidenced that the robustness of complex networks can be comprehensively understood only when considering the strength (weight) of the relationship (link) among nodes^{13,14,15,16,17,18,19,20,21,22}.
The analysis of the robustness of complex weighted networks provided fundamental outcomes. Past studies demonstrated that when the network robustness is measured using the largest connected cluster (LCC), it is highly vulnerable to the removal of links with lower weight (weak links) but robust to deletion of links of higher weight (strong links)^{18,19,20,21}. The widely accepted outcome was that ‘weak links are the universal key for complex networks stability’^{21}. On the other hand, Pajevic and Plenz^{13} outline how the average clustering of nodes (that can be viewed as a measure of the local efficiency of the system) is robust to the removal of weak links but rapidly destroyed when removing links with higher strength. Dall’Asta et al.^{22} showed that introducing the weight of links in the US airports network would decrease its robustness with respect to classic topological frameworks. Further, Bellingeri and Cassi^{23} outlined how the network robustness response to node attacks changes according to the considered measures of the system functioning, i.e. weighted or binary.
In the present work, we investigate the role of the weighted structure of complex networks in shaping their robustness against the removal of both nodes and links. We analyze a high quality set of realworld weighted complex networks from different fields of science (Table 1). The considered network present different number of nodes, links, and a sound interpretation of the nature of link weights, e.g. in the US airports network, the weight identifies the passengers flowing from two airports; in the neural network of the nematode C. Elegans, it identifies the number of connections between two neurons. We randomly removed nodes or links to simulate an error in the system, and we eliminated nodes with higher number of links and with higher strength (e.g. higher sum of link weights) and links with higher weight to simulate an attack. This is the socalled attack strategy with nodes or links removed according to some structural properties of the network^{1,2,3,4,5,6,7,8,9}. We evaluated the robustness of the complex networks to removal of nodes or links in terms of the decrease of network functioning measures reflecting both the binarytopological and the complex weighted structure of the system^{3,7,8,14}. We therefore used the largest connected cluster in the network (LCC)^{1,2,3} and the unweighted efficiency (Eff_{bin})^{14,15,23}, to get a binary (unweighted) measures of the network functioning and we used the efficiency (Eff)^{14,15,23} and the total flow in the system (TF)^{17} to have a weighted measures for the functioning of the networks (Table 2).
Results and Discussion
When links or nodes are removed from the network we can assess the decrease of the system functioning according to different measures as showed in Fig. 1a. The more important are the components removed from the network, the steeper is the decrease in the network functioning measure. For example, in Fig. 1a the red removal strategy identifies more important components in the network, since a given fraction q of nodeslinks removed, triggers a steeper decrease in the network functioning efficiency (Eff, normalized on the initial maximal value) with respect to the black strategy. To compare the response among networks and measures, we resume the removal outcomes in a single value defined as the network robustness (R), reported in Fig. 1b. The value of R corresponds to the area below the curve of the system functioning against the fraction of nodeslinks removed and ranging between two theoretical extremes, R ≃ 0 and R ≃ 1. In Fig. 1 we show the robustness outcomes under different types of nodeslinks removals.
Links removal
When quantifying the system functioning with Eff we find the realworld complex networks to be highly vulnerable to the removal of links with higher weight (Fig. 1c); i.e. Strong strategy produces the fastest decrease of the system efficiency functioning (Eff). Further, we found realworld complex networks Eff to be very robust to the deletion of weaker links and Weak links removal strategy is highly ineffective even causing an efficiency decrease lower than the random removal of links (Rand). At the opposite, when measuring the network functioning with LCC, we find Strong and Weak links attack strategies the less and the most effective to reduce the LCC, especially for Cargo and Air networks (Fig. 1c). Only in the social networks (Net and UK), the deletion of strong links is able to vanish the LCC faster than Weak links removal. This confirms recent analyses showing that the deletion of strong links preserves the LCC until the weakest links are removed^{18,19,20,21}, except for the realworld social networks under study.
In realworld networks, link weights are coupled to the binary topology in a nontrivial way^{18,19,20,21,22}, for instance with nodes strengthdegree correlation meaning that links with higher weight are more likely joining high degree nodes^{16,17}. For this reason it is important to understand if strong links are important in supporting Eff because of their large weight or of their specific occurrence among more connected nodes (hubs) of the network. We compute the robustness of the realworld networks after weights randomization over the topological structure, with link weights independent of any topological features and acting as a control outcome. Strong strategy results to be the most harmful to decrease Eff even when randomizing the real weights over the binarytopological structure, i.e. the importance of strong links to support Eff is maintained also when the real weighttopology coupling is not correlated (Fig. 1c, see Eff_{ran}). We then perform an experimentum crucis by removing links according to the real link weights, then measuring Eff_{bin} considering links like binary (all weights equals to 1). In this way we nullify the influence of the weight to shape the efficiency (Eff) and we maintain only the binarytopological role of the strong links. We find an efficiency reversal pattern for all the realworld complex networks where weak links removal (Weak) readily decreases the efficiency functioning whereas Strong becomes ineffective (Fig. 2a, Eff − Eff_{bin} column). In other words, the importance of strong links to support the information delivery efficiency is mainly due to their larger intensity, with a secondary role of their topological positioning. These results bring important evidences inside the long standing debate about the importance of weak and strong links^{13,18,19,20,21,24} showing that links carrying larger weight would be fundamental to support the efficiency of the system hence not being the main responsible of the topological connectedness of the network. We also revise the importance of weak links in support network robustness confirming their function in maintaining the topological connectedness of the network^{18,19,20,21} but also their small relevance to the information delivery efficiency (Figs 1c, 2a). Very important, we outline that removing a small fraction of strong links can readily reduce the realworld network information delivery efficiency (Eff) despite the size of the largest connected cluster (LCC) is still preserved. In other terms, the binary network functioning parameters Eff_{bin} and LCC are constant under the removal of links with higher weight (strong links) whereas Eff and TF experience a fast decrease and leave the system in a ‘connected but inefficient’ network state. For example, removing a small fraction of busy shipping routes (10% of the strong links) from Cargoship network produced a quick collapse in the system efficiency (losing 50% Eff) while isolating only 2% of portsnodes (2% LCC) (Fig. S8, Strong row). Further, in C. Elegans network the removals of 10% of strong links induced the 60% Eff decrease yet leaving roughly unaltered the size of the LCC (Fig. S8, Strong row). This discover outline an underappreciated but most likely pattern of network failures: namely that most failures may occur even though the network is completely connected. For the sake of an example, we illustrate this finding in Fig. 3a–c for a subgraph of the US Airports network. After the removals of 16 links of highest weight, i.e. 10% of the links (Fig. 3b), the system loses 29% of the total flow (TF, black bar) and the 35% of the system functioning efficiency (Eff, red bar) but only the 4% of the binary efficiency (Eff_{bin}, blue bar) and no decrease in the largest connected cluster (LCC, green bar). After the removals of 26 links of highest weight (17% of the links), the total flow (TF) is the 44% of the initial flowing meaning that the removal of only 17% of the links is able to roughly halves the number of passengers from this subset of airports (Fig. 3c). Further, the system functioning efficiency (Eff) collapses to 38% of its initial efficiency. At this final step where all the 26 strongest links are removed, the network loses only the 12% of the binary efficiency (Eff_{bin}) and only one node is disconnected from the network with a decrease in the largest connected cluster (LCC) of the 4%.
Nodes removal
We find that realworld networks are robust to the random removal but vulnerable to the deletion of higher connected and higher strength nodes for all the functioning measures, i.e. realworld weighted networks would be “random resistance” and “attack prone” (Fig. 1d) confirming classical binarytopological outcomes^{1,2,3,4,5,6,7,8,9}. Very interesting, removing a few nodes abruptly collapse the functioning efficiency (Eff) of the networks (Fig. 4). For example, in Fig. 3d–f we depict the outcomes of the removal of the three highest strength nodes in the C. Elegans neuronal network. Removing the first highest strength node triggers the 27% of the efficiency (Eff) and the 20% of the total flow TF decrease, but only 3% of the LCC and 2% of the Eff_{bin}. (Fig. 3e). Removing the three main nodes the C. Elegans is deprived of the links with higher weight (dark grey links) with a sharp decrease of the efficiency Eff (50%) and of the total flow TF (32%). Nonetheless, a multitude of weak interactions (soft grey links) holds the network still connected showing a minimal LCC (4%) and Eff_{bin} (5%) decrease (Fig. 3f). We find the same system vulnerability in the other realword networks, for example removing 5 nodesmetabolites among the 1100 total nodesmetabolites in E. Coli network decreases the 30% of the efficiency and only the 2% the LCC (Fig. 4); in the US Airports network, 5 nodesairports removals over the N = 500 airports sharply decreases the total flow (TF) to the 60% of the initial value with 5% LCC decrease (Fig. S10); further, 5 nodesports attack reduces to the 70% the efficiency in the Cargoship network (Fig. 4). Only for the Human brain network we find a small difference in the measurements with Eff close to the LCC (Figs 4 and S10). In all the networks the binary efficiency (Eff_{bin}) follows the LCC with a slow decrease. These outcomes can be resumed in a novel and problematic response pattern of the realworld complex networks, where the removal of nodeslinks playing a major role in the energyinformation delivery may leave the system in a connected but inefficient state (Fig. 3).
Dall’Asta et al.^{22} showed that when removing highly connected nodes the total ‘outreach’ (e.g. a measure computing the product the link weight and the Euclidean distance covered by the link connecting airports) of the US Airports network decreased much rapidly than its LCC measure. Our findings wide Dall’Asta et al.^{22} outcomes for different kinds of network and measures able to evaluate its weighted structure, unveiling that even removing a few nodes can leave realworld complex weighted networks in a connected (the LCC and Eff_{bin} are preserved) but inefficient state (Eff and TF quickly collapse). Comparing the normalized LCC and Eff trends under few nodes removals, we observe narrow difference in the network functioning measurement, with a sharper decrease of the network efficiency Eff. This difference is quantified by the pink area between the curves in Fig. 4 for all the realworld complex networks. For example, for the Net scientist coautorship network, after 5 higher strength nodes removals we find Eff collapsing below the 50% of the initial value where instead the LCC is around the 85% (Fig. 4, Str column). In the Coli network, 5 high degree nodes removals only, are able to decrease Eff below the 75% leaving the LCC roughly unchanged (Fig. 4, First column). These evidences outline how using binary measurements like the widely used largest connected cluster (LCC) may overestimate the robustness of realworld networks.
Very interesting, we find that realworld networks exhibit higher efficiency (Eff) robustness to the removal of nodes after the link weights randomization (except for the Human brain) (Fig. 2b, Eff − Eff_{ran}, and Fig. S7). We find an important higher efficiency (Eff) robustness to nodes attack of the randomized networks for a few of nodes attack as well (Fig. S11). In networks with strengthdegree correlation, the removal of higher connected nodes (the socalled hubs) will delete strongest links with higher information loss; differently, in the control network, the weights randomization eliminates the correlation and the removal of hubs would intercept less strong links with a reduced decrease in the information delivery efficiency. This finding indicates that some level of nodes degreestrength coupling discovered in realworld complex networks^{16,17,22,24} would make these systems even more vulnerable to nodes attack.
The role of weights heterogeneity
To examine in depth how link weights pattern influences the robustness of realworld networks we removed nodeslinks increasing the link weights heterogeneity, by assigning link weights sorted from a rectangular and from a two values distribution (Supplemental Material S2). The link weights heterogeneity is enhanced by progressively increasing the maximum weight (Wmax). We find that the robustness of the network efficiency (Eff) is negatively correlated with link weights heterogeneity; differently, the robustness of the largest connected cluster (LCC) remains roughly constant by increasing the link weight heterogeneity (Fig. 5). This discovery would indicate that the overestimation of the network robustness adopting the LCC measure may be higher in realworld complex networks owing larger heterogeneity in the weights of the links.
Surprisingly, increasing Wmax we assist to a decrease in the robustness (Eff) for all the nodes removal strategies (Fig. 6a) whereas the proportion of total flow (TF) subtracted to the networks remains roughly constant (Fig. 6a). For random links removal we find the same pattern (Fig. 6b). All these trends are more pronounced for the two values distribution of link weights (Fig. S11). This indicating that the decrease in robustness efficiency is not due to a major amount of ‘weight’ intercepted in the networks with increasing Wmax, but it would be an effect of the larger link weights heterogeneity (See Supplemental Material S3 for additional results and discussion). Differently, the total flow subtracted in the network by Strong and Weak links removal is clearly related to the weights heterogeneity, and we observe a robustness decrease for strongest (and an increase for weakest) links removal for both TF and Eff measures (Fig. 6b).
This last discovery of the negative relationship between network efficiency decrease and link weights heterogeneity shows that the robustness decrease is not only related to the transition from binary (LCC) to weighted measurement as suggested from Dall’Asta et al.^{22}, but it is within a more general mechanism by which enhancing link weights heterogeneity negatively affects the robustness (Eff) in realworld complex networks (Figs 6, S12). The heterogeneity in link weights is interpreted as a feature able to stabilize different realworld networks^{21,25} with these systems selforganizing toward large heterogeneity in link weights (with many weak links). Our discoveries suggest that if the realworld complex networks are systems evolving toward larger link weights heterogeneity, they may pay the price in terms of robustness, with potential higher vulnerability to nodeslinks failure.
Conclusion
The problem of measuring the robustness of a network to targeted attacks or to random failures of its nodeslinks components is rapidly becoming major topics of research in the complex network community^{26}. This for the urgency to reduce the adverse consequences associated with realworld systems failures, and especially of critical infrastructures. The robustness of complex networks is mainly investigate focusing on the binary version of these systems. Our study showed that adopting measures of functioning able to evaluate the link weights structure strongly change the response of the realworld weighted networks under attack or failure. We find that the attack of a very small fraction of nodeslinks can trigger an abrupt collapse of the weighted functioning measures while parameter that evaluate the simple binary connectedness are almost unaffected, i.e. the attack toward few nodeslinks returns these systems in what we called a connected but inefficient state. This finding has important implications since the most of the realworld system failures may occur even though the network is completely connected and for this reason, adopting a binary measures of the network functioning like widely used the largest connected cluster (LCC) may overestimate the network robustness.
On the other hand, to investigate the reasons bearing the robustness of realworld networks is the gate to understand the mechanisms underlying evolution of these complex systems. Most of the studies focusing the binary structure of the realworld networks showed yet classical outcomes about the evolution of realworld networks, such as their smallworld characteristic^{1,7}, e.g. nodes are close each other in realworld networks, or their robustness to failure in spite of higher levels of vulnerability to the attack of higher connected nodes^{1,2,3}. Analyses of weighted networks showed that links of lower weight play a key role in sustaining the connectivity of the system outlining the importance of weak links for realworld complex networks stability^{21}. In this contribution we furnished a complementary perspective outlining the importance of links with higher weight for the realworld network functioning, by showing that a small fraction of strong links removal rapidly decreases the efficiency and the total flow in these systems.
Last, we investigated how the difference in the weight of the links affects the robustness of realworld networks. We find that the artificial increase of the link weights heterogeneity produced a faster efficiency (Eff) decrease under attack and failures. This finding open a new perspective showing that realworld networks with higher level of weight heterogeneity associated to the links may be less robust to attacks or failures.
Our study leads to several additional observations and interesting directions for future investigations. First, to analyze the robustness of realworld networks under attacks where nodeslinks are ranked according to weighted properties, such as the weighted betweenness centrality, the weighted closeness centrality or others. To understand which attack based on weighted properties of the nodes is able to trigger the higher damage in the network will be a novel approach in the research investigating the nodes attack strategies efficacy. Second, to encourage studies exploring what we consider an important but underappreciated fact of network failures: namely that most failures occur even though the realworld networks are completely connected. This can be the case of many daily life examples in transportation networks, such as a street closure in a road network or the personal strike closing nodesairports in the flying network. But it can be also a frequent occurrence in biological and social networks, like the inhibition of some nodesneurons functioning in brain networks or the leaving of individuals in working community. In all these examples from different reality domains, the network functioning is expected to decrease even though its general connectivity is still preserved. Last, we showed an interesting negative correlation with link weights heterogeneity and the robustness of realworld networks. Since realworld weighted networks owns significant level of heterogeneity in links weights^{21,24,27,28} our last result bring important implications for the robustness of the real systems. This result comes from numerical simulations based on seven given real networks. We hope it could inspire future theoretical investigations that might bring to more general mathematical proves of the observed relationship.
Methods
Nodes removal strategies
We follow 3 nodes removal strategies:

Rand: nodes are randomly removed. This type of removal represents the possibility of failure (error) in the network^{8,9}. For the random process of nodes removal (Rand) we averaged the outcomes from 10^{4} simulations.

First: nodes are removed according to the degree of the nodes, i.e. the number of links to the node^{8,9}. The number of the links to the node corresponds to the neighbors nodes number. First is the binary nodes removal strategy, not considering the weight of the links and it represents the classical network attack^{8}.

Str: nodes are removed according to the strength, i.e. the sum of the weight of the links to the node^{23}. This is the weighted counterpart of the binary attack strategy First.
Links removal strategies
We follow 3 links removal strategies:

Rand: links are randomly removed. This type of removal represents the possibility of links failure (error) in the network^{13}. For the random process of links removal (Rand) we averaged the outcomes from 10^{4} simulations.

Strong: links are removed in decreasing order of weight, i.e. links with higher weight are removed first^{13} and it represents an attack directed to the links.

Weak: links are removed in increasing order of weight, i.e. links with lower weight are removed first^{13}.
The sequential removal of nodeslinks is evaluated on the initial configuration of the network (i.e. before any removals) and not recalculated. In presence of ties (i.e. nodes or links with equals ranking properties) nodeslinks with equals rank are randomly sorted.
The measures of network functioning
The largest connected cluster (LCC)
The largest connected cluster (LCC) is a widely used measure of the network functioning^{1,2,3}. The LCC is also known as giant component and it is the highest number of connected nodes in the network. Complex network may present different clusters, e.g. different subsets of nodes connected among them but disjointed with nodes belonging to other clusters. Let be S_{j} the size of the jth cluster in the network, the LCC is the maximum cluster size:
The removal of nodeslinks may disconnect nodes belonging to the giant component thus producing the LCC measure decrease. The LCC decrease indicates that if we start from one node chose at random in the largest connected cluster, less nodes in the network are reachable. The LCC is a simple indicator evaluating the topological connectedness of the network and it not account the weight of the links that join nodes belonging to the largest connected cluster. We can see the LCC as an indicator of the possibility to reach nodes in the network with no consideration about the magnitude of the linkspathways we have to travel. For this reasons we adopt it like a binary measure of the network functioning not reflecting the heterogeneity of the link weights.
The total flow (TF)
The total flow is the sum of link weights and it represents the actual or the potential flowing between nodes pairs in the network. For example, in the transportation network of the US Airports the TF measure represent the actual flows between nodes pairs (where ‘actual’ means the flying passengers in a year); also in the transportation Cargo ship network TF represents the actual flow indicating the shipping journeys between ports in a year. Differently, in the Human brain network, where the weight of the link indicates the fiber connection density among brain regions, TF that is the sum of all the weights is a potential energyinformation flowing along the fibers connecting nodes pairs, or in the C. Elegans realworld complex weighted network, TF indicates the total number of connections realized between pairs of neurons. In other terms, TF can be viewed as the thermodynamics capacity or a quantity influencing the actual flow between nodes pairs in the network but do not uniquely determine it, e.g. the higher is the connection density in the brain network, the higher can be the information delivered between brain regions in the unit of time.
Let be the weighted network G_{w} can be represented by a N × N matrix W where elements w_{ij} > 0 if there is a links of weight w between nodes i and j, w_{ij} = 0 otherwise.
The total flow is computed as:
The total flow in the system (TF) is the sum of the link weights not reflecting the global weighted structure of network. In particular, TF not consider the links paths necessary to travel among nodes or the related global information delivery efficiency (like the Eff measures) or the nodes connectedness (as the LCC). Thus, when the network is subjected to the removal of nodes or links in our simulations, TF will not evaluate the following elongation of the paths among nodes or the nodes disconnections (or other combined changes) in the global network structure, but only it accounts the decrease in the total weight of the links. Since TF does not evaluate the global network structure we can see it like the simplest measure evaluating the weighted networks functioning.
The weighted efficiency (Eff) and the binary efficiency (Eff _{bin})
The concept of efficiency of the network was proposed by Latora and Marchiori^{14} with the aim to introduce a measurement encompassing the difference in link weights in the evaluation of the weighted networks functioning. The efficiency of a network is a measure of how efficiently it exchanges information. On a global scale, i.e. considering all the nodescomponents of the system, the efficiency quantifies the exchange of information across the whole network where information is concurrently exchanged. The efficiency is a robust and widely used measure of the network functioning^{14,15,23,29}.
The efficiency measurement is based on the shortest paths notion^{17}. In network science, the shortest path is the path between two nodes in a network such that the sum of the weights of its constituent links is minimized. In a binary network a shortest path between a couple of nodes is an integer number computing the minimum number of links to travel from one node to the other. To compute the shortest path in a weighted network we have to account the difference in link weights and to do this we first compute the inverse of the link weights. This is a standard procedure^{14,15,23,29} with the aim to decrease the length of the links with higher weight and increase the length of the links with lower weight. In this way nodes joined by links with higher weight are more close to each other. This procedure has a straight meaning in realworld weighted networks. For example, in the C. Elegans network, two nodesneurons joined by higher number of linksconnections (higher weight) are closer than couples of neurons with lower number of connections (lower weight). For this reasons we can see the inverse of the link weights as a ‘weighted distance’ covered by the link and higher weight links as ‘shorter and faster routes’ for the information travelling. Thus, in a weighted network the shortest path between a couple of nodes is the minimum sum of the ‘weighted distances’ necessary to travel from the nodes. In the Supplemental Materials S1 we furnish a step by step example of the shortest paths computation in binary and weighted networks.
The efficiency of the network is then defined:
where N denotes the total nodes in a network and d(i, j) denotes the length of the shortest path between a node i and another node j. When we include weight to the links (computing the weighted shortest paths) we obtain the weighted efficiency Eff; in the case we do not associate weight to the links (computing the simple binary shortest paths) we obtain the binary efficiency Eff_{bin}. For a detailed explanation of the network efficiency measurement see Supplemental Materials S1. The efficiency here defined works with binary or weighted networks as well but it furnishes different interpretation of the information delivery in the system. Eff_{bin} considers the information delivery as all the network links owns equal capacity; differently, Eff precisely evaluates the effect of link weights heterogeneity to shape the shortest paths in the network and thus affecting the information delivery efficiency. We point out that the efficiency (Eff) and the total flow (TF) are measures able to account the heterogeneity in link weights with different meaning. The efficiency (Eff) is measure based on the shortest paths evaluating the global network structure capacity to deliver information among nodes and a decrease in the efficiency indicates a reduction in the energyinformation pace exchanging over the network. The total flow (TF) is the sum of the link weights not accounting the shortest paths and the global network structure; thus a decrease in the total flow (TF) indicates a simple reduction in the link weights sum over the network. For example, supposing we remove some strong links in the networks, the Eff measure will evaluate both the higher weight lost in the system and the elongation of the shortest paths, whereas the TF will only evaluate the simple reduction in the link weights sum.
Realworld complex networks with modified link weights
We modify the weight of the links maintaining the binary topological structure of these systems. We first randomly assign to each link a weight sorted from the rectangular distribution tuning the sole parameter Wmax indicating the allowed maximum links weight. Then we randomly assign the weight to the links sorting from the bimodal distribution where links weight can have only two distinct values with equals probability (1, Wmax). For both the distributions, we tune Wmax with values {1, 2, 3, 5, 7, 10, 20, 50, 100, 1000, 100000}; for Wmax = 1 all the links have weight = 1 (higher network homogeneity) and the network is binary; for Wmax ≥ 2 the network starts to exhibit variance in link weights with the max variance for Wmax = 100000 (higher network heterogeneity). In this way the binary topological structure of the realworld network is maintained but any pattern of correlation between the topological and the weighted structure is eliminated, making possible to understand how the simple heterogeneity in links weights affects the system robustness. Since the link weights assignment is a random procedure, we perform 10^{4} simulations for each Wmax value.
References
Albert, R., Jeong, H. & Barabasi, A. Error and attack tolerance of complex networks. Nature 406, 378–82 (2000).
Callaway, D. S., Newman, M. E., Strogatz, S. H. & Watts, D. J. Network robustness and fragility: percolation on random graphs. Phys. Rev. Lett. 85, 5468–71 (2000).
Albert, R. & Barabási, A. Statistical mechanics of complex networks. Rev. Mod. Phys. 74 (2002).
Cohen, R., Erez, K., benAvraham, D. & Havlin, S. Breakdown of the Internet under Intentional Attack. Phys. Rev. Lett. 86, 3682–3685 (2001).
Holme, P., Kim, B. J., Yoon, C. N. & Han, S. K. Attack vulnerability of complex networks. Phys. Rev. E 65, 056109 (2002).
Crucitti, P., Latora, V. & Marchiori, M. Model for cascading failures in complex networks. Phys. Rev. E 69, 045104 (2004).
Crucitti, P., Latora, V., Marchiori, M. & Rapisarda, A. Error and attack tolerance of complex networks. Phys. A Stat. Mech. its Appl. 340, 388–394 (2004).
Bellingeri, M., Cassi, D. & Vincenzi, S. Efficiency of attack strategies on complex model and realworld networks. Phys. A Stat. Mech. its Appl. 414, 174–180 (2014).
Iyer, S., Killingback, T., Sundaram, B. & Wang, Z. Attack robustness and centrality of complex networks. PLoS One 8, e59613 (2013).
Bellingeri, M. V. & Bodini, A. Threshold extinction in food webs. Theoretical ecology 6, 143–152 (2013).
Motter, A. E., Gulbahce, N., Almaas, E. & Barabási, A. L. Predicting synthetic rescues in metabolic networks. Mol. Syst. Biol. 4 (2008).
Bellingeri, M., Agliari, E. & Cassi, D. Optimization strategies with resource scarcity: from immunization of networks to the traveling salesman problem. Mod. Phys. Lett. B (2015).
Pajevic, S. & Plenz, D. The organization of strong links in complex networks. Nat. Phys. 8, 429–436 (2012).
Latora, V. & Marchiori, M. Efficient Behavior of SmallWorld Networks. Phys. Rev. Lett. 87, 198701 (2001).
Latora, V. & Marchiori, M. Economic smallworld behavior in weighted networks. Eur. Phys. J. B 32, 249–263 (2003).
Barrat, A., Barthélemy, M., PastorSatorras, R. & Vespignani, A. The architecture of complex weighted networks. Proc. Natl. Acad. Sci. USA 101, 3747–3752 (2004).
Boccaletti, S., Latora, V., Moreno, Y., Chavez, M. & Hwang, D. U. Complex networks: Structure and dynamics. Phys. Rep. 424, 175–308 (2006).
Kumpula, J. M., Onnela, J. P., Saramäki, J., Kaski, K. & Kertész, J. Emergence of communities in weighted networks. Phys. Rev. Lett. 99 (2007).
Garas, A., Argyrakis, P. & Havlin, S. The structural role of weak and strong links in a financial market network. Eur. Phys. J. B 63, 265–271 (2008).
Garlaschelli, D. The weighted random graph model. New J. Phys. 11, 073005 (2009).
Csermely, P. Weak Links. Front. Collect. 421, https://doi.org/10.1007/9783540311577 (2009).
Dall’Asta, L., Barrat, A., Barthélemy, M. & Vespignani, A. Vulnerability of weighted networks. J. Stat. Mech. Theory Exp. 04006 (2006).
Bellingeri, M. & Cassi, D. Robustness of weighted networks. Phys. A Stat. Mech. its Appl. 489, 47–55 (2018).
Allard, A., Serrano, M. Á., GarcíaPérez, G. & Boguñá, M. The geometric nature of weights in real complex networks. Nat. Commun. 8, 1–8 (2017).
Granovetter, M. The Strength of Weak Ties. Am. J. Sociol. 78, 1360–1380 (1973).
Wandelt, S., Sun, X., Feng, D., Zanin, M. & Havlin, S. A comparative analysis of approaches to networkdismantling. Scientific Reports 8, 13513 (2018).
Colizza, V., PastorSatorras, R. & Vespignani, A. Reactiondiffusion processes and metapopulation models in heterogeneous networks. Nat. Phys. 3, 276–282 (2007).
Serrano, M. Á., Boguñá, M. & Sagués, F. Uncovering the hidden geometry behind metabolic networks. Mol. Biosyst. 8, 843–850 (2012).
Bellingeri, M., Bevacqua, D., Scotognella, F., LU, Z. M. & Cassi, D. Efficacy of local attack strategies on the Beijing road complex weighted network. Phys. A Stat. Mech. its Appl. 510, 316–328 (2018).
Watts, D. J. & Strogatz, S. H. Collective dynamics of ‘smallworld’ networks. Nature 393, 440–2 (1998).
Latora, V., Nicosia, V. & Russo, G. Complex Networks: Principles, Methods and Applications. (2017).
AvenaKoenigsberger, A. et al. Using Pareto optimality to explore the topology and dynamics of the human connectome. Philos. Trans. R. Soc. B Biol. Sci. 369 (2014).
Hagmann, P. et al. Mapping the structural core of human cerebral cortex. PLoS Biol. 6, 1479–1493 (2008).
Newman, M. E. J. Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E 74, 036104 (2006).
Nepusz, T., Petroczi, A., Negyessy, L. & Bazso, F. Fuzzy communities and the concept of bridgeness in complex networks. Phys. Rev. E 77, 016107 (2008).
Acknowledgements
Many thanks to E. Agliari for a seminal suggestion on the main idea of this paper. We thank S. Vincenzi for constructive comments during this work. We thank A. Allard and M. Boguna for sharing their realworld networks dataset. We are thankful with two anonymous reviewers which suggestions greatly improved the manuscript.
Author information
Authors and Affiliations
Contributions
Michele Bellingeri, conceived, performed the analyses and wrote the paper; Daniele Bevacqua, wrote the paper; Francesco Scotognella, wrote the paper; Davide Cassi, conceived and wrote the paper.
Corresponding author
Ethics declarations
Competing Interests
The authors declare no competing interests.
Additional information
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Bellingeri, M., Bevacqua, D., Scotognella, F. et al. The heterogeneity in link weights may decrease the robustness of realworld complex weighted networks. Sci Rep 9, 10692 (2019). https://doi.org/10.1038/s41598019471192
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598019471192
This article is cited by

Identifying influential nodes in complex networks using a gravity model based on the Hindex method
Scientific Reports (2023)

Complex network analysis and robustness evaluation of spatial variation of monthly rainfall
Stochastic Environmental Research and Risk Assessment (2023)

Identifying influential spreaders by gravity model considering multicharacteristics of nodes
Scientific Reports (2022)

Identifying influential spreaders in complex networks by an improved gravity model
Scientific Reports (2021)

Paper acceptance prediction at the institutional level based on the combination of individual and network features
Scientometrics (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.