复杂网络科学导论 (8).pdf

《复杂网络科学导论 (8).pdf》由会员分享，可在线阅读，更多相关《复杂网络科学导论 (8).pdf（10页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、ECONOMIC SCIENCESCOMPUTER SCIENCESNetwork effects govern the evolution ofmaritime tradeZuzanna Kosowska-Stamirowskaa,1,2aG eographie-Cit es,CNRS&Universit e Paris 1 Panth eon-Sorbonne,75006 Paris,FranceEdited by Jon Kleinberg,Cornell University,Ithaca,NY,and approved April 15,2020(received for revie

2、w April 24,2019)Maritime transport accounts for over 80%of the world trade vol-ume and is the backbone of the global economy.Global supplychains create a complex network of trade flows.The structure ofthis network impacts not only the socioeconomic developmentof the concerned regions but also their

3、ecosystems.The move-ments of ships are a considerable source of CO2emissions andcontribute to climate change.In the wake of the announced devel-opment of Arctic shipping,the need to understand the behaviorof the maritime trade network and to predict future trade flowsbecomes pressing.We use a unique

4、 database of daily movementsof the world fleet over the period 19772008 and apply machinelearning techniques on network data to develop models for pre-dicting the opening of new shipping lines and for forecastingtrade volume on links.We find that the evolution of this sys-tem is governed by a simple

5、 rule from network science,relyingon the number of common neighbors between pairs of ports.This finding is consistent over all three decades of temporal data.We further confirm it with a natural experiment,involving traf-fic redirection from the port of Kobe after the 1995 earthquake.Our forecasting

6、 method enables researchers and industry to easilymodel effects of potential future scenarios at the level of ports,regions,and the world.Our results also indicate that maritimetrade flows follow a form of random walk on the underlying net-work structure of sea connections,highlighting its pivotal r

7、ole inthe development of maritime trade.maritime trade|transport networks|network science|evolving networks|machine learningIn global supply chains,centers of production and consump-tion are far away from each other,creating a complex networkof trade flows.Over 80%of all cargo,in terms of volume,isc

8、arried by sea,accounting for 70%of the total value of inter-national trade(1),with ships being the least expensive means oftransportation in terms of marginal cost per item(2).Maritimetransport is regarded as the backbone of global trade and of theglobal economy(3).Maritime trade flows impact not on

9、ly the economic develop-ment of the concerned regions but also their ecosystems.Movingships are an important vector of spread for bioinvasions(4,5),especially for marine species.At the same time,the futureof the maritime transport industry is inextricably linked to cli-mate change:The movements of s

10、hips contribute significantlyto global CO2emissions(6),and,conversely,future shippingroutes are likely to be affected by the consequences of cli-mate change.With the development of Arctic shipping becominga reality,the need to understand the behavior of the systemof maritime trade flows and to forec

11、ast their future evolutionreasserts itself.Despite the obvious and crucial importance of maritime logis-tics to the world economy,only a few works provide a detailedoverview of the global distribution of maritime trade flows(4),and even fewer analyze their long-term evolution and the ruleswhich gove

12、rn it(7,8).Current research on maritime transport and ports tends tofocus on particular operators or segments of shipping industry,regions,and specific snapshots in time(9 12),whereas stud-ies analyzing historical evolution of the maritime network as awhole are scarce(13).This is presumably due to t

13、he difficultyof development or acquisition of a global temporal database onmaritime trade flows.The only comprehensive statistical sourceallowing for the representation of global maritime trade flowsaccurately and over a long period,developed by the main mar-itime insurer,Lloyds,has only begun to be

14、 exploited,due torestricted access to these sources and lack of adequate technicaltools.In this study,we aim to fill this gap by using data on dailymovement of the world fleet between 1977 and 2008 provided byLloyds List Intelligence.We treat the available data(16.9 mil-lion recorded ship voyages)wi

15、th tools from complex systems andmachine learning on graphs.Our goal is to enable the extractionof economically viable information about trade and trade vol-umes from data that are not only port-specific but also dependon the broader structure of connections between the ports estab-lished by the mov

16、ements of ships.This approach is motivatedby the fact that vessels are functionally comparable to road orrail infrastructure and should be seen as such while designingtransport or trade policies(14).In this paper,we introducetools to model maritime trade flows and simulate the effects ofpotential sh

17、ocks or changes to the system at local,regional,andglobal scale.Network generative models are used to explain mechanisms ofnetwork evolution by assigning probabilities of creation to indi-vidual potential links based on characteristics of the involvednodes.From the point of view of complexity scienc

18、e,genera-tive models for real-world networks have been in the spotlight(15 17)ever since the acclaimed preferential attachment modelwas proposed by Barab asi and Albert(18).Our work develops aSignificanceOver 70%of the total value of international trade is carriedby sea,accounting for 80%of all carg

19、o volumes.Maritimetrade flows impact both the economic development of theconcerned regions and their ecosystems.Shipping routes areconstantly evolving and likely to be affected by the conse-quences of climate change,while,at the same time,ships area considerable source of pollution.This work perform

20、s a rig-orous and comprehensive analysis of maritime trade flows ona global scale over a long period,taking into account aspectsof evolution,reactions to shocks,and discovering predictivemodels.We use machine learning methods to uncover thesingle dominant dynamics that govern the evolution of thestr

21、ucture and the intensity of maritime trade.Author contributions:Z.K.-S.designed research,performed research,analyzed data,andwrote the paper.yCompeting interest statement:Z.K.-S.is affiliated with NavAlgo.yPublished under the PNAS license.yThis article is a PNAS Direct Submission.y1Present address:N

22、avAlgo,91120 Palaiseau,France.y2Email:.yThisarticlecontainssupportinginformationonlineathttps:/www.pnas.org/lookup/suppl/doi:10.1073/pnas.1906670117/-/DCSupplemental.ywww.pnas.org/cgi/doi/10.1073/pnas.1906670117PNAS Latest Articles|1 of 10Downloaded at Akademiska Sjukhuset on May 26,2020 generative

23、model of link opening specifically for the mar-itime trade network.We also apply a similar methodology toexplain maritime trade flow volumes over existing links.Inboth scenarios,this study does so based on temporal,historicalnetwork data.Our research design consists in uncovering the rules governing

24、the evolution of maritime trade network ex post and is almostentirely data driven.The network features,which are the inputvariables fed into the considered learning methods,include botha preselected set of conventional network measures found witha metaoptimization technique and port characteristics

25、suggestedby the literature.We then feed combinations of these featuresinto different learning methods to automatically construct thebest models for predicting link creation and future trade flows.This process produces models that hit the sweet spot betweenprediction accuracy and the amount of data n

26、eeded as modelinput.We believe that the scientific approach proposed in thiswork,which consists in automatic learning of network generativemodels directly from available temporal data(analysis of newlycreated edges and weight changes between two time frames),may also be of independent interest in ot

27、her areas of complexityscience.In this study,we have found that models relying on featuresof the pure network topology,in some cases augmented by addi-tional information on sea distances,are the most powerful forpredicting link creation and estimating trade flows in maritimetrade for all vessel type

28、s.In all contexts,we obtain the most accu-rate forecasts when relying on one specific network feature:thenumber of common neighbors between a pair of ports,under-stood here as the number of other ports that are simultaneouslytrading partners for both of the ports in question(see Fig.1 forillustratio

29、n).We find that the relative quality of models is stableacross the studied period(1977 2008).We have evaluated the performance of models relying onnetwork features against classical gravity models that use dis-tances and data on demographic and economic development todescribe the affinity between po

30、rts,an approach popular in geo-graphic and economic literature(2)and frequently used as abaseline for evaluating the significance of other effects such asthe network connectivity index(19).Our research methodol-ogy encompasses both types of models and automatically tunestheir parameters(where applic

31、able)to their best performance.We have found that classical gravity models perform very poorlyand that,in the case of maritime trade flows,they are very farfrom a“fact of life”(20).Indeed,our results suggest that classi-cal gravity is almost as poor a baseline model for maritime tradeas a hypothesis

32、 of random uniform link creation that requiresno input.We also benchmarked against a tailored variant ofthe gravity model which uses,as weights,port throughputs inthe gravity equation(4),with affinity between ports moderatedthrough their sea distance and measures of cultural and histori-cal ties suc

33、h as sharing a common language,country,or colonialorigin(21),in a combination tuned to optimal performance.We call this model port gravity.We have found port gravitynot to perform well in predicting link creation and to performmoderately well in predicting actual trade flows.It seems thatthese model

34、s miss an important piece of the story:that affin-ity between ports is influenced by network effects which arisefrom preexisting network infrastructure.The nonnetwork vari-ables conventionally used in the gravity models are no more thana partial proxy for some of these effects(SI Appendix,section L)

35、.Armed with this knowledge,we have additionally testedthe effect of common neighbors in a natural experimentthedestruction of the port of Kobe by an earthquake in 1995 and thesubsequent redirection of trade flows forced by this unfortunateevent.We have found that the simple number of common neigh-bo

36、rs shared with Kobe in 1994 successfully identifies ports takingover traffic from Kobe,at least as well as and,in some ranges,better than the models relying on economic data.This connec-tion further supports our claim that network effects,specifically,network-based affinity,govern the evolution of m

37、aritime trade.Looking more generally,our findings support a vision of tradein which units of goods follow a form of random walk on theunderlying network structure.The observed effects provide hintsas to the precise nature of this random walk process which turnsout to be local,as it relies on informa

38、tion accessible only in theports neighborhood.This effect is observed despite the natureof the maritime industry,which involves strong concentration ofcapital and very few global key players controlling most of theworlds shipping market(22).ResultsDataset and Approach.We use a unique,temporal datase

39、t of dailymovements covering the majority of the world fleet(SI Appendix,section A)between 1977 and 2008,developed by the mainmaritime insurer,Lloyds,licensed to the European ResearchCouncil(ERC)“World Seastems”project.The raw database hasbeen cleaned and treated to extract only meaningful movements

40、of ships(SI Appendix,section B).For the learning process,we construct yearly snapshots of themaritime trade network,where ports stand for nodes,and linksare created by ship voyages between two ports in a given year.A link is created as soon as the total vessel dead weight ton-nage(DWT)transported al

41、ong it exceeds a certain threshold(SI Appendix,section C,as well as SI Appendix,section K wherewe consider an alternative network definition for liner shipping).Fig.1.The concept of common neighbors.The two common neighbors(or“common connections”)of the ports of Tangier and Palermo are Marsaxlokk(Ma

42、lta Freeport)and La Spezia,represented by the full red nodes.This example comes from the studied network of container carriers in 2007.2 of 10|www.pnas.org/cgi/doi/10.1073/pnas.1906670117Kosowska-StamirowskaDownloaded at Akademiska Sjukhuset on May 26,2020 ECONOMIC SCIENCESCOMPUTER SCIENCESModeling

43、maritime trade as a network has recently becomea common practice in the industry(23)and is justified by therelative rigidity of sea connections,resulting from both techno-logical reasons(port types,infrastructure,depth,capacity etc.)and business reasons.In our study,the overlap of links appear-ing i

44、n the network in two successive years is in the range of 55 to65%(reference value for model of uniform independent randomconnections:0 to 1%,depending on year and vessel type),withmore-frequented links being preserved more frequently.We construct a separate network for each of the main com-mercial v

45、essel types:container carriers,dry bulk carriers,gen-eral cargo vessels,petroleum tankers,and liquefied natural gastankers(accounting,in total,for 93%of the world fleet DWT inthe database).This division into subnetworks is made because itis known that different vessel types follow different movement

46、patterns(4,13),and some use specialized ports.In the modeling process,we first consider a broad set offeatures(characteristics of ports and the links between them),suggested by the literature on network formation(24 26)andworks on spatial interactions and international trade(2,27,28).The considered

47、port characteristics include demographic andgross domestic product(GDP)economic variables,such as pop-ulation potential and country(SI Appendix,section D)and portthroughput,as well as features that arise from the comparativeplace of the port in the broader network of maritime connec-tions:so-called

48、network centrality measures.Link characteristicsinclude features that arise from relations between the pairs ofports:sea distance,hop distance,number of trade partners incommon(common neighbors measure),or cultural and histor-ical ties.A model F based on port and link features produces,in general,a

49、prediction value wi,jwhich,depending on the con-text,describes the probability of opening a link between a pair ofports,i,j,or describes the flow value between these ports.Thisis written aswij=F(PortFeaturesi,PortFeaturesj,LinkFeaturesij).1The following research design ensures that we discover the b

50、estof the models constructed with the available features.In orderto preselect dominant features which are strongly reflected inthe observed network structure,we apply a metaoptimizationtechnique:symbolic regression(29)on formulae of generativenetwork models for a static snapshot of the network.We th