Mark Azzam

German Aerospace Center–DLR,Koenigswinterer Strasse 522-524,53227 Bonn,Germany

Evolution of airports from a network perspective–An analytical concept

Mark Azzam

German Aerospace Center–DLR,Koenigswinterer Strasse 522-524,53227 Bonn,Germany

Airport classification;Air transportation;Cluster analysis;Complexity science;Network development;Network science

Analyzing airports’role in global air transportation and monitoring their development over time provides an additional perspective on the dynamics of network evolution.In order to understand the different roles airports can play in the network an integrated and multidimensional approach is needed.Therefore,an approach to airport classification through hierarchical clustering considering several parameters from network theory is presented in this paper.By applying a 29 year record of global flight data and calculating the conditional transition probabilities the results are displayed as an evolution graph similar to a discrete-time Markov chain.With this analytical concept the meaning of airports is analyzed from a network perspective and a new airport taxonomy is established.The presented methodology allows tracking the development of airports from certain categories into others over time.Results show that airports of equal classes run through similar stages of development with a limited number of alternatives,indicating clear evolutionary patterns.Apart from giving an overview of the results the paper illustrates the exact data-driven approach and suggests an evaluation scheme.The methodology can help the public and industry sector to make informed strategy decisions when it comes to air transportation infrastructure.

1.Introduction

As early as 1945 Caroline and Walter Isard foresaw that air transportation will have a vital effect on the global market structure leading to a new spatial division.1This statement has never been more applicable than today,since mobility is a characteristic of the new global world order2where a new division of labor3emerged and high-cost,low-bulk freight including perishable commodities is being distributed globally irrespective of its production site.This crucial impact of air transportation on economic activity is commonly considered to occur on two levels.4(1)The ‘enabling or catalytic impact’incorporates economic activity that relies on the‘availability of air transportation services’.(2)The ‘direct,indirect and induced employment impact’refers to ‘the employment in the aviation industry’,the employment along the according supply chain and the employment generated‘by the spending of those directly and indirectly employed’.One can deduce the significance of the two aforementioned levels of economicimpact for the development of a city and region.While an important metropolis is rather dependent on the catalytic impact of air transportation,a city with a major hub-airport particularly benefits from direct,indirect and induced employment.But it also provides accessibility to its hinterland which again is dependent on the catalytic impact.Furthermore,over time cities with hub-airports benefit from their central position in the air transportation network and the strategic decisions made by airlines.New companies that are in search of premises with good global accessibility are likely to locate in the vicinity.These economic mechanisms make clear why public authorities have always focused on infrastructure investments as a means of regional development.But the air transportation network has undergone a great deal of change in the past decades,which can be traced to the growing national economies and major investments in the Far and the Middle East,to innovative aircraft technologies,to market deregulation and to the rise of low cost carries.The network transition made it difficult for public authorities and airport executives to make the right strategic decision for their airport.And according to the Boston Consulting Group5the first step is to soberly assess the airports’rolein thenew network to determinepossible investments and carrier strategies.

The objective of this paper is to contribute to the assessment of airports and their changing role within the global air transportation network.The paper presents a new approach to classifying airports from a worldwide network perspective and monitoring their development over time to reveal evolutionary patterns.There are some studies that analyze the meaning of airports to the global air transportation system.Many of these use a theoretical traf fic model to analyze traf fic flows and the airports’relevance to the system.6–8Others try to identify structures within the real data traf fic flow9or apply measures from network theory.10However,none of these studies actually describe airports and their meaning to the overall system from a complex network perspective.The different approaches presented in these studies also do not cope well with a large set of airports or longitudinal data.

The issues of airport classi fication and airport evolution from a network perspective have been addressed in very few scienti fic studies.The closest reference to this paper can probably be found in Malighetti et al.11and Burghouwt and Hakfoort12but there are other studies with a similar focus.13,14However,these studies mainly focus on the hub-qualities of airports in certain world areas and lack the global view.In addition,these studies do not aim to analyze the airports’changing role over time but try to determine a current state.While the approach presented in these studies helps to analyze a large set of airports in their relation to the overall network they do not present a method to analyze airport development.The substantially new methodological approach presented in this paper allows classifying airports and revealing their changing role in the network over time.It also finds general patterns of airport evolution and reasons them with their geographical context.Last but not least,this study is unique in its temporal and geographical scope,due to the exceptional 29 year global flight data record.The presented approach combines network theory,cluster analysis and probability theory in the form of an evolution graph inspired by the Markov chain.The results are validated by adapting a known evaluation model and presented by a new visualization scheme for better interpretation.This comprehensive data mining approach was developed specifically for this research study to cope with the extremely large data set that was available for analysis.The methodology is specified in the following.

2.General approach

Clustering is the methodological basis for the analytical concept presented here.Cluster analysis as a computer-based pattern recognition tool allows revealing structures and regularities within data.15In contrast to supervised classification where the classes are known beforehand,unsupervised classification or clustering is a data driven approach that groups all tuples with similar features.16The algorithms are designed in a way that the objects within one identified group have maximum similarity among each other and minimum similarity to objects of other groups.17The analytical concept presented here allows identifying meaningful airport classes based on various parameters from network theory.The applied data set is the Official Airline Guide(OAG)flight schedules database from 1979 to 2007.The cluster analysis is not performed repeatedly on a yearly basis,but instead the airport data for the entire time period is clustered at once and the tuples merely keep their time stamp for subsequent analyses.This means,that the applied data set consists of a large number of tuples,where each tuple represents different network parameters for a single airport at a certain time.This approach has the crucial advantage,that new classes of airports that have evolved later in time can be identified as such.

Fig.1 Evolution graph.

Fig.2 Schematic view of the analytical concept.

3.Methodology

There are two fundamentally different techniques for clustering.18In partitioning relocation clustering such as K-means the number of clusters is determined beforehand.Then based on an initial seed assignment the single objects or tuples are reassigned to the clusters in an iteratve optimization process until the objective function converges.19Thus,the process is not deterministic.Hierarchical clustering is applied in this research especially because it is a deterministic approach and promises good success in the process of identifying a reasonable number of clusters.The algorithm only needs to be run once to identify the clustered data structure for all possible numbers of clusters.In addition,this method of clustering brings forth a hierarchy of clusters,in which every cluster consists of smaller clusters and is part of a superordinate cluster.There is a great advantage to this flexibility in the degree of clustering and granularity,especially for its application to airport clustering.18The downside is,that this non-iterative approach leads to less optimal cluster assignments.However,there is a way to optimize the assignment of tuples to the clusters at the end similar to the partitioning clustering,as shown later.In hierarchical clustering there are two different procedures to build the hierarchy.The agglomerative algorithm starts out with every data object or tuples representing a single cluster before merging them step by step into larger clusters.The divisive algorithm starts out with all tuples merged in a single cluster before partitioning them step by step.20Due to an advantage in computational speed for comparable results the agglomerative algorithm is preferred.16

3.1.Feature extraction and selection

Feature extraction and selection is decisive for the research results when performing a cluster analysis.It is important that only those features,that provide unique information and are relevant to defining the classes,are included in the analysis.21The aim is to provide maximum relevant information in the data while keeping the number of features to a minimum.Otherwise the number of identified classes easily exceed a sensible limit in terms of interpretation.

Due to the objective of classifying airports according to their role in the global air transportation network the applied features in this study draw on parameters from network theory.A comprehensive introduction to network theory is given in Boccaletti et al.22In the context of this research the air transportation network is described as a weighted digraph,where the airports are represented as vertices and direct flights are represented as edges.Summer and winter flight plans were furthermore aggregated to yearly data,which represent the single time steps.The applied parameters contribute to three different categories of characterization:the general scale of the airport,its utilization and its integration into the global air transportation network.From all considered network parameters the following were chosen for clustering based on extensive preliminary testing within this study.

Seat weighted strength r(s):Is the sum of all seats on scheduled direct flights to and from the airport.This feature characterizes the required infrastructural airport capacity on the terminal side for airport logistics such as baggage handling.In addition it gives information about the potential economic turnover of an airport.23

Frequency weighted strength r(f):Is the sum of all inbound and outbound flights.This feature characterizes the required infrastructural airside capacity of an airport such as the runways capacity.Combined with r(s)it determines the average aircraft capacity operating at the airport.

Degree j:The airport degree is the sum of directly connected inbound and outbound origin-destination pairs,therefore,itisavaluableparameterin termsofnetwork integration and utilization.

Accessibility*a:Accessibility and closeness centrality,respectively,is the degree to which the airport is connected to rest of the network.It quantifies the average number of flight segments on the shortest paths to all other airports.Therefore,it also describes the degree of global network integration.The better accessible an airport is,the more likely it is that the region will benefit from the catalytic impact of air travel.(*:To increase the validity of this feature for the applied yearly aggregated network data,it was calculated based on the remaining giant component of a network that was filtered so that only connections with a service frequency≥360 remained.)

Betweenness centrality$b:This parameter quantifies the number of shortest paths between any pair of airports in the network that connects via the airport.Therefore,it is a very important feature in terms of utilization and network integration.It allows evaluating the potential hub quality of an airport.Airports that have a high centrality lie along the shortest paths between many airport pairs and therefore necessarily have a geographical advantage to offer onward connecting travel.($:To increase the validity of this feature for the applied yearly aggregated network data,it was calculated based on the remaining giant component of a network that was filtered so that only connections with a service frequency≥360 remained.In addition every compatible path was considered as shortest path.Compatibility was assumed where a path had a maximum detour of 20%to the actual shortest path.)

3.2.Preprocessing the data

The analyses are based on the OAG flight schedules from 1979 to 2007.The network was filtered in different ways beforehand.Only flights that had a yearly flight frequency of≥52 were taken into account.And only airports that had at least one direct connection operated with a frequency of?≥360 p.a.were considered sufficiently ‘active’.In addition,for running the cluster algorithm a sample was drawn from the data for computational reasons.Since the study focusses on rather relevant commercial airports the sample was not drawn randomly.Of all airports in the filtered Network only 1094 have been active(in the defined sense)over the entire time period.Of these only the 585 largest that accounted for over 95%of the scheduled flights in 2007 are taken into account when performing the cluster analysis.To further reduce the sample size,the sample only considers 15 time steps between 1979 and 2007,one every 2 years:079,081,...,005,007.The remaining data tuples of active airports and all time steps are assigned to the identified clusters subsequently.Hence,of the 32,759 tuples(airports considered active at different time steps)only 8775 are structured by the cluster algorithm,while the remaining are assigned afterwards according to their best fit.

After calculating the network parameters from the filtered network data they have to be scaled to a fixed range.Otherwise their impact on the clustering might be disproportionate.Features with high variance and magnitude would dominate those features with lower variance and magnitude.That is because the cluster algorithm applies the same metric to every feature when identifying similarities.With the exception of accessibility a all features are rescaled linearly24to the interval[0,100].q(q:The foundation for this and the following scaling procedure is that the features are at least of interval scale.However,the error for features with ordinal scale is usually too little to justify the effort of applying a special procedure25)Accessibility is first rescaled using a softmax normalization and then rescaled to the closed interval[0,100].Softmax normalization uses a sigmoidal function for rescaling so that outlier data is transformed within the saturation region while central values are rescaled rather linearly in a more sensitive way,Fig.3.Therefore,the procedure is very robust against outliers.26The method was chosen for the parameter accessibility to increase the impact of small variance between values on the lower end where most airports lie,while reducing the impact on the higher end where very few small airports had very bad accessibility values.

Fig.3 Softmax normalization.

3.3.Proximity measure

The two core elements of agglomerative hierarchical clustering are the proximity measure and linkage rule.Both together determine how the homogenous groups are formed,meaning which data objects are similar and in which order or hierarchy they are grouped together.The proximity measure determines how similar two objects are based on distance.The linkage rule defines how the proximity measure is applied to groups of objects.

Choosing the right proximity measure mainly depends on the parameter’s scale of measure.The Minkowski distance that is derived from theLP-norm is most commonly used.The distance d between two data objects(X,Y)is then calculated according to Eq.(1).27

Hence,the larger the value p the more weight is given to the largest differences between single features of two data objects.For features of interval or ratio scale usually the Euclidean distance that follows fromp=2 is applied.Euclidean distance is in line with natural human perception of distance and therefore leads to similar results as a visual classification.17Due to the considered features in this research the Euclidean norm was chosen.

3.4.Linkage rule

The first steps of agglomerative clustering merely cluster single data objects to pairs,therefore,the proximity measure alone determines similarity.As soon as tuples are grouped together,there are different ways to apply the proximity measure.Similarity can for example be defined as the sum of distances between all data objects of two groups or the average of distances.The linkage rule defines in which manner the proximity measure is applied to groups of tuples.19,28,29Based on extensive preliminary testing in this study centroid linkage(also known as UPGMC(unweighted pair group method using centroids))proved most suitable in the context of this work.The similarity between two groups is defined by the distance between the two respective centers of the tuples.This approach is limited to features of at least interval scale.The advantages are that it is computationally undemanding and can be applied to large datasets.It is very robust and does not give preference to clusters of globular shape but can identify clusters of all shapes instead.It also has a tendency to generate clusters with similar variances among their features.17

3.5.Number of clusters

The approach of identifying a reasonable number of clusters aims at obtaining clusters that are well separated,while penalizing an increasing number of clusters.TheL-Method30was chosen here due to its efficiency and good performance for hierarchical clustering algorithms.This Method is an approach to automatically identify the point of maximum curvature,also referred to as ‘knee’,in an evaluation graph.The evaluation graph also called ‘structogram’is a two dimensional plot with the number of clusters on the abscissa and the similarity measure on the ordinate.Fig.4 displays the well-known dendrogram on the left and the structogram on the right.The knee of the graph can be interpreted as the point of transition from a high to low gain in cluster separation for an increasing number of clusters.Thus the knee indicates the number of clusters with an ideal trade-off between cluster separation and number of clusters.The algorithm identifies the knee by fitting two linear functions to all pairs of sequences of data points on the left and right side of each potential knee.The knee lies where the two linear regression functions have minimal residuals.

There are several other procedures to identify the ideal number of clusters such as gap statistics31or the CH-index.32Since it is understood that airport classes cannot be absolutely distinct,there are probably a range of good sets of classes and knowledge based evaluation necessary.For this circumstance the L-method proved to be more powerful than the gap statistic,CH-index or several other methods.31,32

For the analysis presented in this paper theL-method suggested an ideal number of clusters of eight,indicated by rhombus in Fig.4.However,as the structogram has a very smooth course without a distinct knee,there is not only one good result.Consequently,the chosen number of clusters depends on the desired granularity for further investigations.An in-depth analysis of the last four cluster mergers(from 12 to 8)reveals that here different compact clusters of the same evolution path are aggregated,indicated by hollow triangles in Fig.4.For interpretability of the results it was therefore decided not to adapt these last mergers and instead to keep the 12 separate clusters as an ideal number for further investigations.This action is supported by the cluster evaluation coefficient introduced below.

Fig.4 Dendrogram and structogram.

3.6.Clustering evaluation

Cluster analysis always aims to identify group-structures in data.However,an issue with this data-driven approach is that it will always structure the data into groups,whether or not there are any meaningful hidden group structures to reveal.That is why evaluating the identified clustering is indispensable.33Apart from most important knowledge-based checks for plausibility and meaningfulness through the user,there are also quantitative methods.These performance indicators can usually also help deciding on the best clustering framework like proximity measure or linkage rule through trial and error.

One impartial valuation standard are the so called silhouettes.34This internal valuation technique reveals how central or decentral each data object lies within a cluster.In this way it evaluates how compact and well separated the clusters are.The silhouette values(i)of an objectiis calculated according to Eq.(2).Wherea(i)is the average proximity to all other objects of the same cluster andb(i)is the average proximity to all objects of the closest neighbor cluster to objecti.

4.Results

4.1.Airport taxonomy

The twelve airport classes can themselves be categorized to three archetypes.The first archetype covers the classes No.12 and No.6,which are smaller and undifferentiated airports.These airports often function as less important feeder airports in economically well evolved regions or as important airports in economically weak or extremely remote regions.The second archetype covers the classes No.9,No.8,No.5 and No.4.These are typical point-to-point or origin-destination(O&D)airports.Airports of these classes can be very important to the overall network,but are not very central.That is why they do not serve as hubs primarily.Often they connect important remote economic metropolises in otherwise economically weak or sparsely populated regions to transnational markets.The third archetype covers all types of hub airports.Due to a geographical advantage,a sound regional economy and the strategic decision of at least one airline to pick the airport as their hub,they are often very central in the network in terms of accessibility and betweenness.A detailed overview of all identified airport classes is given below.

Table 1 Characteristics of airport classes indicated by mean values.

Regional Airport I–No.12:This category contains the smallest of commercial airports that primarily regional and narrow body aircraft head for.These airports are generally weakly connected and lie in the network periphery,which separates them to all other airports by 5.6 flight legs on average.Usually operations are highly dominated by single airlines which use them as a regional feeder to their closest hubs.

Regional Airport II–No.6:This category contains the next more important undifferentiated airports.Apart from many regional jets they are also approached by larger narrow body aircraft.With more direct connections and a less regional focus the airports’connectivities are significantly better although on a low level since they still have to be ascribed to the network periphery.However,these airports already attract a variety of airlines which indicates that they serve major cities or popular holiday resorts.

Major Intercontinental O&D Airport I–No.9:Even if the airports of this category do not belong to the largest ones nowadays,they are relatively well connected to the global network.Just about 3.8 flight segments separate them from all otherairportswhich isjustonemorethan thebestconnected in the world.These airports connect important remote metropolises like Melbourne(MEL)and popular holiday resorts like the Maldives(MLE)to some of the most important cities and best-connected airports worldwide.This statement is also supported by the high fraction of intercontinental connections and the large number of strong airlines that operate a mixture of narrow body and wide body aircraft at these airports.According to their weak betweenness centrality they are obviously not embedded in a strong economic region that they could serve as a hub.

Major Intercontinental O&D Airport II–No.8:This category is quite similar to category No.9.However,the two airports Tokyo Narita(NRT)and Singapore Changi(SIN)in this exclusive category belong to the most important worldwide.No other group is as internationally oriented,according to the very large fraction of intercontinental connections and wide body aircraft operations.Although the number of compatible shortest paths connections leading via these airports is very low,the accessibility to the global network is excellent.That these airports serve world cities can also be seen by the large number or airlines that have a high traffic share in this obviously lucrative market.

Major Regional O&D Airport I–No.5:This airport category can be described as a regional complementary counterweight to category No.9.These airports belong to major capitals that already have one large international airport and they developed as a secondary airport for regional traffic.Examples for this development are Seoul with the airports Gimpo(GMP–No.5)and Incheon(ICN–No.2)or Sao Paolo with the airports Congonhas(CGH–No.5)and Guarulhos(GRU–No.9).Although the overall number of scheduled seats is higher than in category No.9 the mixture of operating aircraft is quite similar.However,these airports have a strong regional focus,which puts them in the periphery of the global network with limited network accessibility.They usually connect important cities within an otherwise economically developing world region.Although these airports are dominated by their respective home carrier their betweenness is still quite low.

Major Regional O&D Airport II–No.4:The unique airport of this category pretty much plays the same role as airports of category No.5 except on a larger scale.Tokyo Haneda(HND)is one of the busiest airports in the world and merely serves the regional transportation market,since its counterpart Tokyo Narita(NRT–No.8)offers the intercontinental connections.Interestingly the number of wide body aircraft that operates on short haul flights is enormous.In time other airports like Gimpo or Congonhas could follow Haneda in their development.

Major International Airport–No.1:This category covers the large part of important international airports like Lisbon(LIS)that are well connected to the rest of the network but do not play an extraordinary role or cannot be described as exceptional in any way.As the starting point for airports to become global hubs,this category covers airports that usually connect regional to global air transport markets.That is why this category can mainly be found in world regions with a pronounced region wide economy.In world regions like Asia,South America or occasionally Afrika this category is not featured before the late nineties,e.g.Manila(MNL).The range of operating aircraft spreads from regional jets to wide bodies.The limited market power of home carriers indicates that airports in this category often serve commercial centers,with substantial transregional demand for traffic.Even today airports of this category serve as important hubs.

Major International Hub–No.2:This category features very important hubs with a central role in the global network.Apart from a large number of scheduled seats these airports have a very good direct and indirect connectivity and their betweenness centrality is only exceeded by very few.According to hierarchical clustering two different subcategories can be found within this cluster.The first group contains airports that became a hub because of the strategic decision of an airline to build its network around this airport,e.g.Dubai(DXB).These airports have a lower average distance to all destinations(1483 km)and a higher betweenness centrality(6277).At the same time only 1.3 airlines account for two-third of all scheduled seats.The second group contains major international hubs that are located at global economic centers and do not merely depend on the strategic decision of single hub carriers,e.g.Bangkok(BKK)or Rome(FCO).On average 4.5 carriers account for two-third of the scheduled passengers and the average distance to all destinations is doubled to 2809 km while the betweenness centrality is only 4172.This leads to the assumption that a higher fraction of airport traffic originates by genuine local demand instead of airline driven routing.

Fig.5 Evolution of airports.

International Mega Fortress Hub I–No.3:In 1979 only Hartsfield-Jackson Atlanta(ATL)and Chicago O’Hare(ORD)were assigned to this category.In terms of passengers and flights both belonged to the largest and best connected in the world and both were highly dominated by their home carriers Delta Airlines and United Airlines,who themselves belonged to the largest carriers in the world.That is why the airports are termed fortress hubs.By 2007 several similar airports had followed into this category.With a betweenness of 10,947 only the succeeding category No.7 offers better hub qualities.The different features including the high fraction of narrow body jets and high frequency connections indicate that airports of this category first of all owe their development to the strategic decisions made by the home carriers instead of to the genuine local demand of the metropolitan area.However,these airports are a crucial part of the global network.

International Mega Fortress Hub II–No.7:This airport category logically succeeds category No.3 and comprises ATL since 1999 and ORD since 2007 as its only representatives.The category stands for the busiest,most central and best connected airports in the world.Only category No.11 has a slightly better accessibility to the global network.However,the fraction of continental traffic is quite high and since the single major home carrier accounts for two-third of scheduled seats these airports primarily do not owe their role in the global network to the significance of the metropolitan area they serve but their geographic position and the network-strategic decision of the airlines.

International Mega All-Round Hub I–No.11:Although the airports in this category are significantly less trafficked than those of No.7 they have the best network accessibility in the world.On average only 2.75 flight legs lie between them and any other airport in the world.In addition,the many directly connected destinations are distributed with a stronger intercontinental focus.For the global connectedness of the overall network their high betweenness centrality cannot be valuated high enough.Airports of this category are located at geographically beneficial locations and some of the most important world cities like New York(JFK)or Paris(CDG).Therefore,traffic is generated by genuine local demand which leads to strong competition between the many operating airlines at these airports.

International Mega All-Round Hub II–No.10:The other group of International Mega Allround Hubs is as busy as airports of category No.3.However,unlike fortress hubs the airports in this category are located at some of the most important world cities like London(LHR),Los Angeles(LAX)or Beijing(PEK)and do not profit from the network strategy of single hub airlines.Quite the contrary,no other group of airports has as many airlines being responsible for two-third of the schedules seats.Therefore,these airports are extremely well connected directly and indirectly.Given the number of direct connections and the multitude of operating airlines the betweenness centrality is remarkable.

4.2.Airport development

According to the analytical concept the final step is to create the evolution graph with respect to the airport categories and transition probabilities,to identify evolutionary patterns(Fig.5).Every airport category is displayed in a radar chart with its mean values of the different features as well as the 5%and 95%percentiles.In addition to the twelve identified categories the status of inactivity is also displayed(No.0).The considerable conditional transition probabilities from one category into another are included as edge weights.For simplification,these probabilities do not take into account that in all categories the highest probability is self-referring.Which means that from one time step to another most of the airports remain in the same category.Here only category changes are considered.Starting from the undifferentiated regional airports I/II(No.12/No.6)four developmental paths can be identified,which can be assigned to the remaining two archetypes.Two paths lead to the regional and intercontinental O&D airports.The other two yield the classical hub airports with their differentiation between fortress and all round hubs.

The intercontinental O&D path consisting of the categories No.9 and No.8 has its root in category No.6.A geographical analysis reveals that most often these airports evolve where they connect an important remote commercial capital or popular tourist site in a region that otherwise has no developed air transport market due to the missing economic power in that region.There are some exceptions to this finding.In these cases the airports connect commercial capitals in newly industrialized or emerging regions and are complemented by an airport of the regional O&D path(No.5 or No.6),e.g.Seoul,Sao Paolo or Tokio.This constellation is quite rare and rather unfavorable from a passenger perspective since onward connections from regional to intercontinental always make an airport transfer necessary.

The other two developmental paths lead to the two main hub airport categories all-round hub and fortress hub.These airports integrate well established regional air transport markets with the global network.Therefore,they are often found in highly industrialized world areas.Starting from the category regional airports II(No.6)the airports develop into major international airports(No.1)before turning into international hubs(No.2).In this category it is often already clear,which development an airport takes depending on the affiliation with one of the two subcategories.While one path leads to the international mega fortress hubs I/II(No.3/No.7)the other yields international mega all-round hubs I/II(No.11/No.10).Airports that mainly owe their great development to the strategic decision made by a large network carrier and a convenient geographic position rather develop into fortress hubs.Airports that are located at important world cities or metropolitan areas with a significant global-economic relevance are more likely to become all-round hubs.

5.Conclusions

Analyzing airports’role within global air transportation and monitoring their development over time provides a new perspective on the dynamics of network evolution.For this reason the paper presents a new methodological approach to characterizing airports with coefficients from network theory and establishing a new airport taxonomy.Based on this datadriven method a large set of airports can easily be classified and their development and role within the network monitored over a long period of time without being limited to a small number of case studies.By drawing on hierarchical clustering and determining conditional transition probabilities from the panel data the results are displayed by an evolution graph similar to a discrete-time Markov chain.By taking six parameters from network theory into account,twelve different airport categories are identified that spread along clear evolutionary patterns.The results show that airports with a similar function in the network run through similar development stages.Apart from giving an overview of the results the paper also illustrates the exact methodology and suggests an evaluation scheme.Although the analytical concept presented is not suited for airport forecasting,the results do provide guidance to potential developments and can help the public and industry sector to make informed strategy decisions when it comes to air transportation infrastructure.Further research is needed to interpret the findings from a geo-political and geo-economical perspective.A further study should also focus on detailing the larger categories of small rather undifferentiated airports and also extend the study to recent developments.Especially in East Asia and the Middle East some major developments are expected to be seen.

Acknowledgment

The foundation of this research was supported by the German Research Foundation through the graduate school 1343 as well as the former European Center for Aviation Development-ECAD GmbH.

1.Isard C,Isard W.Economic implications of aircraft.Q J Econ1945;59(2):145–69.

2.Urry J.Mobilities.Cambridge:Polity Press;2007.

3.Krings BJ.Global restructuring and its effects on occupations:towards a new division of labor?Proceedings of the international conference‘technologies of globalization’;2008 October 30–31;Darmstadt,Germany.2008.p.117–30.

4.Ishutkina M,Hansman RJ.Analysis of interaction between air transportation and economic activity.Mass Inst Technol2011;38(5):185–8.

5.BCG.Airports–dawn of a new era:Preparing for one of the industry’s biggest shake-ups(technical report).Boston:The Boston Consulting Group;2004.

6.Burns MC,Cladera JR,Bergad MM.The spatial implications of the functional proximity deriving from air passanger flows between European metropolitan urban regions.GeoJournal2008;71(1):37–52.

7.Matsumoto H.International urban systems and air passenger and cargo flows:some calculations.J Air Transport Manage2004;10(4):241–9.

8.Matsumoto H.International air network structures and air traffic density of world cities.Transport Res Part E:Logist Transport Rev2007;43(3):269–82.

9.Grubesic TH,Matisziw TC,Zook MA.Global airline networks and nodal regions.GeoJournal2008;71(1):53–66.

10.Bowen J.Network change,deregulation,and access in the global airline industry.Econ Geogr2002;78(4):425–39.

11.Malighetti P,Paleari S,Redondi R.Airport classification and functionality within the European network.Prob Perspect Manage2009;7(1):183–96.

12.Burghouwt G,Hakfoort J.The evolution of the European aviation network,1990–1998.J Air Transp Manage2001;7(5):311–8.

13.Adikariwattage V,de Barros AG,Wirasinghe SC,Ruwanpura J.Airport classification criteria based on passenger characteristics and terminal size.J Air Transp Manage2012;24(24):36–41.

14.Rodriguez-Deniz H,Suau-Sanchez P,Voltes-Dorta A.Classifying airports according to their hub dimensions:an application to the US domestic network.J Transp Geogr2013;33(33):188–95.

15.Bishop CM.Pattern recognition and machine learning.Berlin:Springer Science+Business Media,LLC;2006.

16.Marques de Sa JP.Pattern recognition:concepts,methods and applications.Berlin:Springer;2001.

17.Larose DT.Discovering knowledge in data:an introduction to data mining.New York:John Wiley&Sons Inc.;2005.

18.Berkhin P.A survey of clustering data mining techniques.Grouping multidimensional data:recent advances in clustering.Berlin:Springer;2006.

19.Duda RO,Hart PE,Stork DG.Pattern classification.2nd ed.New York:Wiley-Interscience;2000.

20.Witten IH,Frank E,Hall MA.Data mining:practical machine learning tools and techniques.3rd ed.San Francisco:Morgan Kaufmann;2011.

21.Milligan GW.Clustering and classification.Singapore:World Scientific Publishing Company;1996.p.341–76.

22.Boccaletti S,Latora V,Moreno Y,Chavez M,Hwang DU.Complex networks:Structure and dynamics.PhysRep2006;424:175–308.

23.Mensen H.Planung,anlage und betrieb von flugplaetzen.Berlin:Springer;2007.

24.Masters T.Practical neural network recipes in C++.Manhattan:Academic Press;1993.

25.Briand L,El Eman K,Morasca S.On the application of measurement theory in software engineering.Empirical Softw Eng1996;1(1):61–88.

26.Ross A,Nandakumar K.Fusion,score-level.In:Li SZ,Jain AK,editors.Encyclopedia of biometrics.Berlin:Springer;2009.

27.Jain AK,Murty MN,Flynn PJ.Data clustering:a review.ACM Comput Surv1999;31(3):264–323.

28.Theodoridis S,Koutoumbas K.Pattern recognition.2nd ed.Manhattan:Academic Press;2003.

29.Everitt BS,Landau S,Leese M,Stahl DD.Cluster analysis.New York:Wiley;2011.

30.Salvador S,Chan P.Determining the number of clusters/segments in hierarchical clustering/segmentation algorithms.IEEE International Conference on Tools with Artificial Intelligence;2004 November 15-17;Piscataway(NJ):IEEE Press;2004.p.576–84.31.Tibshirani R,Guenther W,Trevor H.Estimating the number of clusters in a data set via the gap statistic.J Roy Statist Soc:Series B(Stat Methodol)2001;63(2):411–23.

32.Calinski T,Harabasz J.A dendrite method for cluster analysis.Commun Stat1974;3(1):1–27.

33.Aldenderfer MS,Blashfield RK.Cluster analysis.London:Sage Publications Inc.;1995.

34.Rousseeuw PJ.Silhouettes:A graphical aid to the interpretation and validation of cluster analysis.J Comput Appl Math1987;20(20):53–65.

35.Kaufman L,Rousseeuw PJ.Finding groups in data:An introduction to cluster analysis.New York:Wiley;1990.

13 June 2016;revised 10 October 2016;accepted 22 December 2016

Available online 14 February 2017

E-mail address:mark.azzam@dlr.de

Peer review under responsibility of Editorial Committee of CJA.