1. Introduction
Railway networks are complex systems that are vulnerable to disruptions caused by natural disasters, human errors, and deliberate attacks. Given the critical role played by railway networks in transporting people and goods, ensuring their resilience is of utmost importance. In general, resilience refers to a system’s ability to cope with, adapt to, and recover from disruptions, disturbances, and failures, while maintaining essential functions. The resilience assessment can be conducted through various methods, including mathematical optimization, data-driven approaches, and network science-based quantification. Moving on to the latter framework, many measures have been proposed for assessing resilience by looking at the network topology. Some of these measures include degree centrality, betweenness centrality, clustering coefficient, and the average path length, which provide insights into different aspects of network structures and functions. Such metrics present some limitations because they only consider the shortest connections between nodes within the network, without taking into account potential alternative routes or modes of transport. Moving in such a framework, this paper proposes an assessment of railway network resilience by means of a local robustness measure, called the effective resistance centrality () of an edge or a vertex, which is able to capture the effect of either a specific vertex or a specific edge on the network’s robustness. In fact, both stations and tracks can be put out of service due to natural, technical, or social hazards, and the use of allows us to provide a comprehensive importance measure method in order to assess the relevance of railway lines and stations. Since the degree to which a networked system continues to function typically depends on the integrity of the underlying network, the topic of system robustness is usually addressed by analyzing how the network structure changes as vertices (or edges) are removed.
The proposed approach allows for identifying critical vertices (or edges) and it can be useful for planning response plans, such as rehabilitation or restoration of the disrupted links. In this paper, we perform an empirical application and explore the behaviors of
measures by using the railway network of Lombardy, the Italian region with the most diffused rail network in the country. We focus our attention on Lombardy because of the high volume of data and because this region has the highest number of train journeys in Italy. The analysis is based on official rail performance data derived from
Viaggiatreno.it (accessed on 10 May 2023), the Italian official website that presents a live timeline of the railways; to present a complete picture of this reality with the actual number of journeys, the data are related to the period October–November 2020. The data-selection process for the Lombardy railway network involved gathering information from official channels and maps provided by Trenord ([
1]) and Rete Ferroviaria Italiana (RFI) ([
2]). The analysis is based on stations (vertices) and journeys (edges) that are effectively present in the official documents of the regional and national railway transport companies. The selection of the specific vertex under scrutiny is derived from the vertices present on the railway map used, and, therefore, it cannot be defined as an exhaustive representation of the entire regional network (considering all stations) due to the fact that the official network provides a partial representation of all the vertices. Moreover, as explained in a subsequent paragraph of this work, we categorized all of the vertices (following the official Italian classification) and calculated the effective number of journeys (for both directions), for a working day, between the considered vertices. The study’s research questions were formulated based on the rationale mentioned above:
How effective is the proposed ERC measure, compared to classical topological metrics, in identifying critical stations and journeys within a railway network?
What are the most important station and railway line connections that contribute to the overall efficiency and resilience of the Lombardy railway network?
The study shows that the ERC measure can provide useful insights into the robustness of railway networks and it highlights the importance of considering geographical positions in a network analysis. This offers a more comprehensive and integrated approach to measuring the importance of railway network components, filling a research gap in the literature. The main results show how the provided measure enables the identification of critical stations and journeys within the network structure. Additionally, we find consistency with the classification provided by the network manager (as sustained in the report [
3]). Furthermore, the geographical positions of some less relevant stations are critical for the connection of the overall network, and their disconnections can compromise the stability and communication of a significant portion of the Lombardy railway stations. Considering the limitations, it is possible to highlight two factors: the first is the geography aspect, which is not fully representative of the entire set of railway stations of Lombardy (more than 400), and the second is represented by data availability, which limits the concentration of the Milan node in only two stations.
The remainder of this paper is organized as follows. In
Section 2, a literature review is provided. In
Section 3, some preliminaries on the graph theory and an overview of the effective resistance centrality are presented. In
Section 4, the implemented methodology is described. We provide a structural description of the Lombardy railway network in
Section 5. In
Section 6, based on 2020 data, the role and significance that stations play in maintaining the structure of the entire network are investigated. A consistent comparison with several well-known topological measures is provided by assessing the vertex importance in order to show how the proposed measure seems to be more sensitive at capturing the network’s behavior after removing the vertices or edges (topological changes in the structure). Conclusions, further developments, and limitations follow.
2. Literature Review
The vulnerability and resilience of transportation networks, especially railways, have received significant attention in recent literature due to the potential for disruptions and degraded functionality that negatively impact urban areas and associated communities, causing severe economic damages (see, among others, [
4,
5,
6,
7,
8,
9,
10,
11,
12]). Previous research has predominantly focused on the vulnerability of railway networks during incidents, as measured by the susceptibility to incidents that result in a significant reduction in network serviceability [
13]. For an exhaustive literature review, see [
14]. Studies have used the complex network theory to explore the vulnerability of railway networks under topological changes. These studies have proposed a variety of measures to quantitatively assess network topology and evaluate the impacts of removing vertices and edges from the network. Moreover, they compared the pre-disaster values to those in degraded states ([
15,
16,
17,
18]). These measures include centrality measures (e.g., degree, betweenness, closeness), regularity measures (e.g., transitivity, clustering coefficients), efficiency measures (e.g., efficiency, costs), connectivity measures (e.g., connectivity, number of clusters, cluster size), residual closeness, and spectral graph measures. Another measure suitable in this context is the residual closeness, provided in [
19], which captures the results of actions (deleting vertices or edges), even if they are small enough not to disconnect the graph. Moreover, spectral graph measures are exploited in this setting by looking at the connection between the eigenvalues of a matrix associated with a graph and the corresponding structures of a graph (see [
20,
21]). However, there is no general method to measure the importance of railway network components. For instance, in [
22], only edge importance is measured, while only vertex importance is measured in [
13]. Indeed, there is limited research specifically focused on local robustness measures for railway networks. Within this context, the effective resistance centrality has been proposed as a local indicator for assessing robustness in the network analysis. This centrality metric has been shown to be suitable, with respect to classical metrics, such as betweenness and degree centrality, and in terms of its ability to capture the local properties of networks. Specifically, effective resistance centrality can catch the localized impact of the node or edge removal on network resilience, providing a more nuanced understanding of network robustness. Overall, the use of effective resistance centrality as a local indicator has contributed to a deeper analysis of network robustness and the potential impacts of damage or disruptions in specific parts of the network. Furthermore, since
is not based on the shortest paths, such as classical metrics, but takes into account all of the alternative paths present in the network, it could be informative in order to reroute traffic along alternative paths, in case of failure or disruption. The analysis of backup possibilities is crucial in designing and evaluating resilient networks. Therefore, the development of a general method for measuring the importance of individual components of a railway network remains an area of active research. Overall, recent literature has highlighted the critical nature of transportation networks, particularly railways, and the need for ongoing efforts to assess and enhance their resilience in the face of potential disruptions.
3. Theoretical Background
3.1. Some Preliminaries on Graph Theory
Let us firstly recall some standard definitions and results about the graph theory (for more details, see [
23,
24,
25]).
A graph involves a pair of sets , where is the set of vertices and E is the set of edges. Let us consider graphs with fixed-order and fixed-size .
An undirected graph is a graph in which if , then , whereas a directed graph (digraph) is a graph in which each edge (arc) is an ordered pair of vertices. The edge connecting vertices i and j is denoted by . Two vertices are called adjacent if they are connected by an edge. A loop is a degenerate edge of a graph, which joins a vertex to itself (also called a self-loop). Multiple edges are two or more edges connecting the same two vertices i, j. In the sequel, we will not consider self-loops and multiple edges. Moreover, we have a weighted graph when each edge is assigned a weight . Furthermore, the so-called strength of a vertex i, denoted by , is equal to the sum of the weights of the edges adjacent to it.
A simple graph is an unweighted, undirected graph containing no graph loops or multiple edges. A path is a sequence of distinct adjacent vertices from vertex i to vertex j. The distance between i and j is the length of the shortest path joining them when such a path exists, and it is set to , otherwise. A graph G is connected if there is a path between every couple of vertices. A subgraph = of G is a graph, such that and . A particular class of subgraphs is one of the induced subgraphs. A subgraph of G is induced by when implies , for each vertex .
A maximal connected subgraph of is called the connected component of G.
A graph is connected if it has exactly one connected component.
A connected graph G is said to be k-edge-connected if it remains connected whenever fewer than k edges are removed. A connected graph G is said to be k-vertex-connected (or k-connected) if it has more than k vertices and remains connected whenever fewer than k vertices are removed.
A cut vertex is a vertex whose deletion increases the number of components. Similarly, an edge is called a bridge (or cut edge) if its deletion increases the number of components.
Let us denote the degree sequence of G as arranged in a non-increasing order, such that , with being the degree of vertex i.
We can associate a non-negative
n-square matrix
to the graph, representing the adjacency relationships between vertices of
G, which is called the adjacency matrix; the off-diagonal elements
of
are equal to 1 if vertices
i and
j are adjacent, and 0 otherwise; if the graph has self-loops, the corresponding diagonal elements of A are equal to 1. Given the diagonal matrix
of vertex degrees, the Laplacian matrix of
G is defined as
. Considering a graph
G with non-negative edge weights
, we denote by
the weighted adjacency matrix and the weighted Laplacian matrix by
, where
is the diagonal matrix of strengths, with elements
. To summarize, the elements of the weighted Laplacian are given by
Since the weighted Laplacian matrix is symmetric, positive semidefinite, and the rows sum up to 0, this entails that its eigenvalues are real, non-negative, and the smallest eigenvalue is zero. Hence, we can order the eigenvalues and denote them as for , such that .
3.2. Robustness Indicators
As we already mentioned throughout the paper, a very important issue regarding complex networked systems is indeed the robustness of the overall system to the failure of its constituent parts (see [
26]). Recently, a graph measure called the effective graph resistance, also known as the Kirchhoff index, denoted by
(for more details, see [
27,
28,
29]), has received growing interest in network robustness research. In [
30], the authors provide a novel robustness measure called effective resistance centrality, based on the Kirchhoff index. In particular, the effective resistance centrality of a vertex (or an edge) is defined as the relative drop of the Kirchhoff index induced by the deletion of this vertex (edge) from the network. This measure is a local robustness measure that is able to capture the effect of either a specific vertex or a specific edge on the network robustness.
Let us now briefly recall the definitions of the edge-based effective resistance centrality and vertex-based effective resistance centrality, as well as some of their properties.
3.2.1. Edge-Based Effective Resistance Centrality
Let be a k-edge-connected graph (with ) of n vertices and m edges, and be the graph obtained by removing the edge , connecting vertices i and j, from G.
The effective resistance centrality
of the edge
is defined as
It is noteworthy to say that, if G is a 1-edge-connected graph, then the resulting subgraph can be disconnected. In this case, when is a bridge, we have that . displays monotonicity with respect to edge removal.
3.2.2. Vertex-Based Effective Resistance Centrality
Let be a connected graph of n vertices and m edges and be the graph obtained by removing the vertex and all of its related connections from G.
The effective resistance centrality
of the vertex
is defined as
is defined by considering the numerator to be the drop of the normalized Kirchhoff index, denoted by
(see, for more details, see [
29]). The quantity
is not always positive, because it depends on the importance of the specific vertex
in the network. On the one hand, this measure enables the identification of strategic vertices whose failure could significantly impact the network’s resilience. On the other hand, it can also be utilized to identify potential vertices whose removal could enhance the robustness of the network.
We extend the robustness indicators, (
2) and (
3), to the weighted case by replacing the weighted Kirchhoff index and the normalized weighted Kirchhoff index.
6. Numerical Results and Discussion
As previously explained, using the websites
Viaggiatreno.it,
Trenord.it,
Trenitalia.it, it was possible to obtain a complete picture of the journeys between the vertices proposed by the official Trenitalia map. Consequently, the number of journeys to and from a determined vertex was calculated to obtain a complete dataset containing the exact number of journeys between the vertices themselves in a working day in November 2020. Furthermore, all vertices were classified using the parameters by RFI (the Italian railway network; in Italian, this is known as Rete Ferroviaria Italiana) in order to evaluate their importance. For each edge, the exact number of passing trains (in both directions) in one working day in October–November 2020 was calculated. The weighted calculations are, therefore, based on the data obtained from the Trenord and Viaggia Treno sites.
We compute the normalized Kirchhoff index
for the whole network and compare it with the values of
, derived by generating a network with the same characteristics and by using different well-known network models. As reported in
Table 2, we obtain a value of
. It is worth noting that this value is greater than the values of
derived by generating a network with the same characteristics, by applying some well-known simulation models. In fact, we exploit the homogeneous structure of the random graph of the Erdős–Rényi model (see [
46,
47]), the heterogeneous structure of the scale-free graph of the Barabási–Albert model ([
48]), and the small-world graph provided by Watts and Strogatz (see [
49]). In this case, simulated networks show a higher robustness than the Lombardy railway network (note that simulated networks were derived by randomly generating 1000 graphs, with the ER, WS, and BA models, respectively. The number of vertices and edges is kept equal to the number of vertices and edges of the Lombardy railway network).
We now focus on the most relevant values of effective resistance centrality, considering the first ten elements for each category for both unweighted and weighted networks; none of the classical network models can fully describe the entire structure of this transport network.
Table 3 shows the stations with the highest vertex-based
for both unweighted and weighted networks. In this table, we do not consider stations that, if removed, would result in the disconnection of the network. A dedicated focus on these stations is presented in
Table 4.
The first interesting element that emerges from
Table 3 involves the presence of eight stations, which are relevant in both the unweighted and weighted networks.
Firstly, from a transportation point of view, considering these data, we find relevant stations (with a large number of trains and passengers per weekday) and some less relevant railway stations. Beginning with the unweighted side, the first element is the Codogno station; it is categorized as Gold and can be defined as a station of good importance; however, compared to the others present in this table, it is certainly an unexpected result. Checking the graph, however, its strategic position emerges within the regional railway network. In fact, it is a vertex through which several lines pass and, even if it is not a cut vertex, its elimination would create strong connection problems between the two sides of the network. Moving on to the Ponte SP station, we have a silver-level station (medium importance), which is a strategic one for the province to which it belongs, since most of the trains of that quadrant and the three railway lines pass through that station. Subsequently, a station not far from Ponte SP is Bergamo; in third place, this is a platinum railway station in one of the most relevant cities of Lombardy. Moreover, we focus our attention on Rho Fiera and Busto Arsizio, these two vertices are connected to each other, are situated in the northwest of the region, and are strategic for journeys from that part of the region. Their importance is, therefore, a confirmation of the network’s structure; eliminating these vertices would have strong repercussions on the overall rail traffic. Finally, analyzing the vertices of the unweighted network, Albate and Cava Carbonara appear in this classification. The first station is important for the city of Como and the nearby border with Switzerland; the second is a station that three railway lines use to pass through and enter (albeit, with a very low frequency). On the weighted side of the table, the first railway station is Ponte SP; in this station, although the number of trains is not at its maximum level, many routes pass through, making these vertices very important. The same reasoning applies to Bergamo; as already mentioned, it is a platinum-level station, and the number of trains passing through it is important, although the number is not very high. On this side of the table, after Codogno, three relevant vertices with very high weights emerge. In fact, Busto, Rho Fiera, and Milano Rogoredo are three hubs that overlook the city of Milan. Most of the trains coming from the northwest and south of the Milan hub pass through them. Milano Rogoredo, in particular, is the second busiest vertex in the network. This station serves as a hub for numerous long-distance and high-speed services, with trains passing through and stopping at this location. To conclude, in the previous table, two other stations (Brescia and Carnate) are considered relevant for weighted and unweighted sides. Monza (on the weighted side) is a station with a high number of trains and is close to Milan, while Desenzano, an important tourist station, joins the Brescia Verona line.
Table 4 presents the analysis of the stations that, if removed, result in a disconnection of the network for the unweighted and weighted railway networks, respectively.
If these stations are eliminated, the network becomes disconnected. It is possible to check whether the situation is rather similar, considering the weighted and unweighted scenarios. In fact, four of the first five, with different rankings, are the same for both columns (Cremona, Pavia, Piadena, and Milan). Among them, there are some busy vertices from a transportation point of view, such as Milan (regional central hub) and Pavia. Pavia is situated on a line towards Liguria and the port of Genoa. There are two other less important stations (from a transportation perspective) but they are geographically situated in the central part of the graph. In fact, Cremona and Piadena (south of the Lombardy region) are two important hubs of the network, albeit with less rail traffic. To conclude, the focus should be mandatory for Treviglio, which is in the second position on the side of the weighted cut vertices. In fact, this is a station with great importance because it is the final station of two suburban lines to and from Milan; moreover, it is located on the directrix for Verona (and, subsequently, Venice); for these reasons (also due to the number of circulating trains), its position in this ranking is not surprising.
In
Table 5, edge-based
values are presented for both unweighted and weighted networks. In this case, the railroad lines are analyzed. The results expressed in the following table can help us to understand the actual importance of edges through the value of the
index.
Regarding the unweighted network, there are strategic lines (Bergamo–Ponte SP, Busto–Gallarate, Milan–Milan Rogoredo, Novara–Rho Fiera), which can be considered as very relevant for the regional transportation railway network. Without them, there would be heavy repercussions on rail traffic for relevant parts of the regional and suburban lines passing through these edges and vertices. Furthermore, some parts of the network do not seem to be strategically relevant, but they are linked to crucial stations where numerous lines nearly intersect (e.g., Codogno–Cremona, Mantova–Piadena, and Mortara–Novara). The computation of highlights the presence of the station of Treviglio, which appears in the first two considered edges. It is relevant because this station serves as both a junction for several lines and a transit section for trains coming from the east of the region. Subsequently, three edges include the stations already mentioned in previous analyses, although not altogether relevant from the transport point of view, in this analysis, these vertices have assumed high importance. In fact, Bergamo–Ponte SP, Mantova–Piadena, and Codogno–Cremona are crucial points in the analyzed network. Moreover, we observe the significance of the Milan hub and the routes coming from the northwest of the region (Busto–Gallarate and Rho Fiera–Novara), where the daily number of railway journeys is particularly relevant. Finally, let us consider the east side of the network, where the railway station of Desenzano is situated. It is a significant vertex that connects two sections of the graph (Brescia–Desenzano and Desenzano–Verona) and is located on the same east–west route as Treviglio, suggesting the strategic importance of that part of the network.
In
Figure 4, we shed light on the most critical edges, as shown in the aforementioned analysis.
In
Table 6, the stations (both cut vertices and non-cut vertices) ranked with different centrality measures are reported.
It is noteworthy that the most relevant stations, according to the classification provided by Trenitalia, are detected by , while the other measures classify some of them in the bottom ranking. In fact, there are only three “platinum” stations in Lombardy, and they are all included in the top 15 stations by , while betweenness centrality, eigenvector centrality, and residual closeness do not consistently provide clear distinctions among them.
In order to highlight this aspect, in
Figure 5, we report the correlation between the rankings of the above-mentioned measures. It is worth pointing out that
has a weak correlation with the other measures, as shown in
Table 6, while the residual closeness has a positive strong correlation with betweenness centrality, indicating a similar classification of the station importance/centrality. Finally, residual closeness has a negative correlation with eigenvector centrality.
It is noteworthy to highlight that, in this classification, all of the platinum stations in the Lombardy region are presented. Moreover, stations in Bologna and Verona, classified as platinum by Trenitalia, belong to the network, but these stations do not appear in the top ranks. This is explained by the fact that only trains that start or end in Lombardy were considered; therefore, Bologna and Verona, which fall outside of the region, have limited effects on the resilience of the network.
Table 7 summarizes the average values for different centrality measures for the groups identified by Trenitalia classification. It can be seen that the ranking provided by
follows the exact same ordering.
The results that emerge from this analysis can be used for a variety of purposes. One important use of the analysis is to identify topological vulnerabilities in the network, which can help with decisions about where to allocate resources for maintenance and upgrades. This can help reduce the risk of disruptions in the transportation sector and improve overall efficiency and reliability. Another important use of the analysis is to identify potential choke points or critical nodes in the network, which can help with emergency planning and response. By understanding the network’s resilience, emergency responders can more effectively prepare for and respond to disruptions, minimizing the impacts on passengers and freight. Finally, an understanding of network resilience can also contribute to more sustainable and environmentally friendly transportation systems. This is because a more resilient network can lead to reduced energy consumption and emissions, as disruptions and delays can lead to increased fuel consumption and emissions. Overall, the results of the resilience analysis can be a valuable tool for optimizing the performance and reducing the risk of disruptions in the transportation sector. The analysis provides valuable insights that can help decision-makers make informed decisions about maintenance and upgrades, emergency planning, response, and environmental sustainability.