Work in progressContent and data on this page are not final.

Approach

The study reads the TfL network as a directed, weighted graph through two analytical lenses. Information filtering uses the Disparity filter, Global Statistical Significance (GloSS) and a sorted-weight baseline to identify the statistically significant backbone of passenger flows. Stations are grouped into functional travel zones with the Louvain method of community detection, and shortest paths are computed with Dijkstra’s Algorithm.

Two comparison methods assess how stable these results are: Kendall’s tau compares backbone rankings across periods and methods, while Normalized Mutual Information (NMI) compares community partitions. Network performance is summarised by two transport network indicators, efficiency and robustness. Results are explored on interactive maps. Their track geometry was extracted from OpenStreetMap through the Overpass API, each directed edge being cut from its route relation between the projected station positions (the short terminus hops where the track is straight fall back to direct station-to-station segments), and is drawn as splines fitted by least-squares B-spline approximation (Deng and Lin, 2014), offset slightly so that parallel services stay legible. Station layers are rendered with Mapbox GL JS and deck.gl (WebGL), plus RailStreak, a custom GPU shader that animates comet trails along every directed edge to show flow direction and intensity across the day. A few assumptions keep the project feasible: NUMBAT 2019 modelled demand stands in for observed flows, 15-minute slots cover the operating day (05:00–00:30), every edge shares a uniform commercial speed, and community detection runs on an undirected graph.

All analytical maps are also available in a dedicated full-screen view, which provides an expanded canvas for detailed exploration of the backbone, community, and infrastructure layers.

What is TfL

Since 2014, TfL open data has been served through a single Unified API: one common format for every mode, free to registered developers. The API covers journey planning, live arrivals, disruptions, fares, cycle hire and network topology. By 2017, according to the same case study, that meant over 80 feeds, more than 12,000 registered developers and over 600 apps built on them. Stations in this study are keyed by NaPTAN, the UK’s national stop-point reference, which links TfL’s operational data to geographic coordinates.

The NUMBAT project (Network Usage Model for Bus And Tube) is TfL’s most comprehensive rail demand dataset: a typical autumn day for each day type, combining fare-card taps with network and survey inputs and adjusted for disruption. It gives link loads, station boarders and alighters and origin–destination flows in 15-minute slots, from 2016 for the Underground, Overground, Elizabeth line and DLR, which makes it one of the few publicly available network-wide 15-minute flow models for a major European transit system. Its yearly editions, NBT19 to NBT25, are the traffic source of this study; as each describes a typical day, year-to-year changes do not necessarily mirror whole-year demand.

Transport for London makes an ideal test bed for this kind of analysis. Since 2000, TfL has grown alongside the city it serves into one of the most extensive transit systems in the world. The rail network spans five modes. The Underground (Tube) is the oldest and largest: eleven lines radiating from a dense central core into six concentric fare zones. London Overground was rebranded in 2024 into six named services — Lioness, Mildmay, Windrush, Suffragette, Weaver and Liberty — stitching together the orbital suburbs the Tube never reached. The Docklands Light Railway (DLR) connects east London, the Isle of Dogs and Heathrow Airport via Stratford. The Elizabeth line (Crossrail), fully opened in 2022, cuts east–west through a 13-kilometre central tunnel between Paddington and Whitechapel, linking Heathrow in the west to Reading and Shenfield in the east. Finally, the Tram network serves the south London boroughs around Croydon and Wimbledon. Together the five modes cover roughly 753 route kilometres and about 510 stations and stops across Greater London’s 1,572 km², carrying close to 5 million journeys on a typical 2019 weekday. Buses, river services and a cable car complete the wider TfL family, but it is these five rail modes that define the city's rapid transit spine. The full network is captured in the official map.

TfL is also among the best-documented transit systems in the world, largely thanks to a long-standing open data commitment that puts real operational data in the hands of anyone curious enough to dig in. The commitment dates from 2007, ahead of the 2012 Olympic Games, when TfL decided that data funded by the public belonged to the public: rather than build every journey tool itself, it opened its feeds to any app developer able to bring travel information to passengers in better ways and on more devices, with transparency and better products and services for customers as the goal, as Stone and Aravopoulou (2018) document in their case study.

Infrastructure

Chopra et al. (2016) show that structural and functional vulnerability are distinct: a station can be peripheral in the topology and still carry heavy flow. Reading the physical layout on its own, before any flow is added, keeps the two apart, so that what follows can be attributed to demand and not to geometry.

Munoz-Mendez et al. (2018) find the same split in London’s bicycle-sharing network: geographically isolated communities stay stable through the day, while central clusters are volatile. The geographic zones set that expectation for the rail network: outlying lines should form stable pockets, and interchange hubs should blur the boundaries between zones.

The Louvain method is a heuristic: on an identical graph its partition can still move slightly with the order of the nodes, so small differences between years are not always structural.

The infrastructure reading asks what the network’s physical layout looks like as a weighted graph, and its goal is a geographic baseline. Community detection with the Louvain method, applied to the inverse-distance graph, groups stations into geographic proximity zones: every connection is weighted by 1/D, where D is the distance between its two stations, so that geographically close and directly connected stations attract one another most strongly. The zones show how the Tube, the Overground and the DLR are laid out before a single passenger is counted, and the Operations section later compares them with the zones that demand creates. A backbone on distance, built with the filters introduced in Approach, then shows which physical connections are structurally essential.

Edge distance distribution

Tube
0246800.250.50.751051015202530% of connections (313)kmsignif.
Overground
0123400.250.50.751051015202530% of connections (111)kmsignif.
DLR
00.511.522.500.250.50.751051015202530% of connections (47)kmsignif.
  • Distance in km
  • Significance, Disparity
  • Significance, GloSS

Applying the Disparity filter and GloSS to distance has one objective: edge relevance. With D as the weight of each physical connection, the two methods ask whether a connection is long for the stations it joins, not only whether it is long. Their significance, 1 − α (where α, defined in the Disparity and GloSS drop-downs of Approach, is the probability of a weight at least as large under a null model), ranks the connections from the most to the least relevant, and the backbone keeps the most significant ones. It is a reading of the layout, before any passenger is counted.

This is not the same as the idea that the longest connections are the most important. Ranking by absolute distance looks at D alone, so the longest connections come first whatever the stations they join. Significance weighs each D against the other connections of its stations, through their degree and strength: a connection that is short in kilometres can be significant at a station whose other links are shorter still, and a long one can be ordinary at a station where every link is long. The two agree only in part: in the 2025 Tube, Kendall’s τ between D and the significance is +0.41 for Disparity and +0.10 for GloSS, and only 10 of the 31 longest connections are also among the 31 most significant ones under Disparity (7 under GloSS).

Station positions come from NaPTAN (National Public Transport Access Nodes), TfL's authoritative gazetteer of every boarding point in Great Britain, and D is the geodesic distance between those coordinates. The list of stations and links comes from the yearly NUMBAT edition, from 2019 to 2025, so every year has its own graph and its own zones. They diverge because the network changes: the Northern line extension adds Battersea Power Station and Nine Elms in 2021, and Barking Riverside joins the Overground in 2022. The editions also differ in what they list: Battersea Park appears in 2020 and 2021 but not in 2022. From 2022 to 2025 the set of stations is the same. Within a year the geometry does not change, so each year’s partition is computed once and serves as the stable baseline for that year’s demand-weighted zones.

The infrastructure analysis characterises the physical skeleton of the TfL network using distance as its primary KPI — the time-invariant length of each directed connection in kilometres. Unlike traffic, this signal does not vary across periods or day types, so the analysis focuses on structural rather than temporal patterns. Applied to the distance-weighted graph, Disparity and GloSS identify which connections carry anomalously high geographic load relative to their endpoint's local distance distribution, exposing the structural backbone of the physical network.

Community detection (Louvain) operates on the same graph weighted by inverse distance (1/D), so that geographically proximate and directly connected stations attract one another more strongly — the resulting communities are spatial clusters that reflect the network's physical topology and track geometry. These infrastructure partitions serve as the reference against which demand-driven traffic communities are compared in the Operations section through the NMI index.

Users

Barrat et al. (2004) show, for weighted networks, that the hubs by topology and the hubs by flow differ: a station with many connections is not always the one that carries the most passengers. This is why the traffic zones are not expected to repeat the distance zones.

Munoz-Mendez et al. (2018) find that geographically isolated communities stay stable through the day while central clusters are volatile, which is how the comparison across periods is read: stable zones at the edges of the network, changing ones in the centre.

Strano et al. (2015) describe how the Underground shifts betweenness to peripheral endpoints and how orbital lines decentralise it, a pattern to look for on the Overground.

The traffic reading asks how passenger demand redistributes across the physical layout seen in the Infrastructure section. Community detection with the Louvain method, applied to the traffic graph, groups stations that exchange more passengers than a random arrangement would predict, which yields operational service zones: where people actually travel together, shaped by service frequency, capacity and connectivity. Because the zones depend on demand, they can move from one period to the next, and comparing the partitions across the six periods (NMI) tells whether they do. A backbone on traffic, built with the filters introduced in Approach, shows which connections carry the flow that matters.

Edge Traffic distribution

Tube
01M2M3M00.250.50.75105101520253035% of connections (313)passengers per weeksignif.
Overground
0200k400k600k800k00.250.50.751051015202530% of connections (111)passengers per weeksignif.
DLR
0200k400k600k00.250.50.751051015202530% of connections (47)passengers per weeksignif.
  • Traffic (passengers)
  • Significance, Disparity
  • Significance, GloSS

Passenger demand comes from the yearly NUMBAT editions: for each day type, the load on every connection and the entries and exits at every station, by period and in 15-minute slots. Each edition describes a typical autumn day, adjusted by TfL for abnormal events such as industrial action or short closures, so year-to-year changes in the zones combine real changes in demand, the network changes of the Infrastructure section and the choice of that typical day, and they do not necessarily reflect whole-year ridership. The day types themselves change in 2022: Monday–Thursday is replaced by Monday and Tuesday–Thursday, with Friday, Saturday and Sunday unchanged.

The traffic analysis characterises passenger flow through the TfL network across six operational periods — Morning, AM Peak, Midday, PM Peak, Evening, and Late — using two information filtering approaches and a sorted threshold baseline. Disparity and GloSS differ in how significance is assessed: Disparity evaluates each connection against the weight distribution of its own endpoints (a local test), while GloSS combines local and global significance scores. Where both methods agree on a connection's importance, the backbone captures a feature robust to the choice of null model; where they diverge, the difference encodes whether a given flow is locally anomalous or globally rare.

Community detection (Louvain) operates on the same traffic-weighted graph and partitions stations into groups where ridership density within the cluster exceeds what a random arrangement would predict — revealing functional travel zones that may cross administrative or geographic boundaries. The resulting communities are compared to infrastructure-defined partitions in the Operations section through the NMI index.

100%0%
Traffic
05:0010:0015:0020:0000:30

Tube Traffic twt

Each heatmap compares how the connections of the selected mode are ranked at two moments of the selected day type. A cell is the Kendall’s τ between the ranking at the moment on its row and the ranking at the moment on its column, so the matrix shows how stable the ranking is through the day.

Purple (+1) means that the two moments rank the connections in much the same order, grey (0) that their rankings are unrelated, and amber (−1) that the order is reversed. The diagonal compares each moment with itself, so it reads +1. Rows and columns are the periods, or the 15-minute slots in 15-min aggregation.

Blocks of purple mean that the same connections stand out across those moments; purple that fades away from the diagonal means that the ranking drifts through the day; amber means that connections ranked high at one moment rank low at another. The three heatmaps rank the same traffic in three ways, by weight (Sorted), GloSS and Disparity: where they look alike, the result does not depend on the filter.

Rank correlation

Sorted

-1
+1

GloSS

-1
+1

Disparity

-1
+1

Operations

Chopra et al. (2016) show that structural and functional vulnerability are distinct: a network can be sound in its layout and still be strained by the flow it carries. Setting the zones that the layout defines against the zones that demand defines is a way of measuring that distance directly.

Barrat et al. (2004) show, for weighted networks, that the hubs by topology and the hubs by flow differ. The same graph can therefore be partitioned in two different, equally legitimate ways depending on whether its connections are weighted by distance or by passengers.

Munoz-Mendez et al. (2018) find that geographically isolated communities stay stable through the day while central clusters are volatile. If the divergence between the two zonings changes little from one period or day type to another, it comes from how the network is built and used, not from a single peak.

The operational challenge is the distance between where the network is built to connect and where people actually travel. It is measured by comparing two zonings of the same stations: the geographic proximity zones of the Infrastructure section and the operational service zones of the Users section.

For each mode and year, the Louvain method is applied to the same graph under two weightings. With the inverse distance (1/D) as weight, there is a single partition per mode, which describes the layout. With the passengers on each connection as weight, normalised within the period, there is one partition for every day type and period. The normalised mutual information (NMI) between the distance partition and each traffic partition is a number between 0, when the two zonings share no information, and 1, when they are identical, computed over the stations present in both. Each cell of the matrices below is one of these numbers: periods in the rows, day types in the columns, one matrix per mode.

The Louvain method is not fully deterministic: it starts from a random order of the stations, so two runs on the same graph can return slightly different partitions. To make the comparison fair, every partition shown here, for every mode, year, day type, period and weighting, was computed with the same fixed seed on graphs built in the same station order. All of them share one starting point, so a difference between two cells reflects the data and not the randomness of the algorithm. Repeating the runs with other seeds moves the NMI by about 0.01, which is the smallest difference worth reading.

Each cell is the NMI between the geographic zones and the operational zones for one day type and period. The three matrices share one colour scale, from the lowest value of the selected year to 1 (its ends are given under each matrix), so read the number and not only the colour. From 2022, Monday–Thursday is replaced by Monday and Tuesday–Thursday.

A high value means that passengers travel within the zones the layout defines; a low one, that demand crosses their boundaries. A cell apart from its row or column marks a period or day type in which the zones reorganise; similar cells mean that the divergence does not depend on when people travel.

A divergence that persists is the operations gap, reduced by infrastructure expansion or by operational emphasis (frequency, capacity, interchange) on the connections that cross zone boundaries. The matrices show where and when it is wider, not which response suits it.

Community comparison: geographic zones vs operational zones

Tube

0.500
1.000

Overground

0.500
1.000

DLR

0.500
1.000

Beltjens (2020) studies the Paris Métro with a flow-based model: the passenger flows are computed on the network, the backbone of its edges and lines is extracted with the disparity filter, and robustness is assessed by simulating the failure of lines and stations.

The robustness KPI used here follows the same idea of failure simulation, applied to one connection at a time: take it out of service, route its passengers again and measure what they lose. The efficiency KPI describes the service that the connection delivers, so the two are read together: what a connection gives, and what its absence would cost.

Beltjens also shows the disparity filter at work on a real transit network with passenger flow as the weight, which is why the same filter is used here to rank the connections.

The first part of Operations compared two zonings of the network. This part looks at the service that runs on it, through two KPIs calculated for every connection. Robustness asks how much its passengers would lose if the connection stopped working, and efficiency asks how much passenger service it delivers for its length and for the trains that run on it. Together they tell apart a connection that delivers a lot of service but is easy to replace from one that delivers less but has no alternative.

Both KPIs are calculated for every connection, period and day type, ranked with the information filtering introduced in Approach, and their rankings are compared across periods in the analysis below. They are available only for NUMBAT 2019, the edition for which origin–destination matrices exist. Open a KPI to see its formulation, processing and algorithm.

The Sorted threshold method was selected as the reference ranking for the robustness indicator. Because the robustness KPI is itself derived from the shortest-path algorithm — which encodes the network's structural topology through betweenness centrality — the resulting edge weights already carry an intrinsic topological signal. A simple sorted rank preserves this structure without imposing an additional statistical model, producing a direct, parameter-free representation of the network's vulnerability hierarchy. For the efficiency indicator, Disparity and GloSS were preferred because efficiency is fundamentally a traffic-based KPI — composed of observed passenger flows weighted by distance and speed. Both information filtering methods are designed to detect statistically significant flows against a null model, making them better suited than a plain sorted rank at capturing the demand-driven backbone. The heatmaps below show the Spearman rank correlation between the Sorted robustness ranking and both the Disparity and GloSS efficiency rankings across all operational intervals. Periods of high positive correlation indicate that structurally critical edges and the most efficient traffic connections coincide — concentrating operational pressure on the same links. Periods of low or negative correlation reveal a beneficial structural separation: the connections most exposed to disruption are not those bearing the greatest functional load.

Information filtering
Sorted
100%0%
Period

Tube Robustness Mon-to-Thu

Conclusions

The three networks analysed — Underground, Overground and DLR — differ substantially in scale, topology and ridership, which makes them a natural testbed for evaluating analytical methods across different levels of complexity. A method that performs well on the relatively compact DLR graph must also remain interpretable when applied to the denser, more interconnected tube network, and the range of outcomes across the three systems reveals both the strengths and the boundary conditions of each approach.

Among the backbone filtering methods considered, neither the disparity filter nor GloSS can be declared categorically superior — each preserves a different notion of relevance, and the choice between them is governed by the analytical objective rather than by raw performance. What the comparison does establish clearly is the relative inadequacy of pure weight ranking as a filtering strategy: retaining edges solely on the basis of absolute traffic volume systematically discards the locally significant connections that give the network its structural coherence. The graph-theoretic methods are, in that specific regard, unambiguously better suited to the task of delineating a faithful traffic-flow blueprint.

The traffic analysis exposes a consistent directionality in passenger flow that shifts predictably across the day. Morning periods are characterised by strong inbound movement toward central interchanges, while evening periods reverse this pattern into dispersed outbound flows. This inversion is most pronounced on weekdays, where the AM and PM peaks are sharply defined; on weekends the directional asymmetry softens considerably, replaced by a more diffuse and temporally spread demand that reflects leisure rather than commute patterns. The distinction has practical implications for any capacity or resilience assessment that uses a single representative day as its input.

Community detection applied to the network yields complementary readings depending on the edge weight used. When weights are drawn from physical connection distances, the resulting communities reflect the geography of infrastructure: clusters correspond to corridor groups, branch families and engineering boundaries that define how the network was built. When weights are drawn from passenger traffic, the communities shift to reflect user displacement patterns — groupings that are shaped by where people actually travel rather than by how track was laid. Effective network operation seeks to minimise the divergence between these two readings: a system whose infrastructure communities closely mirror its demand communities is one whose capacity is well-matched to its riders.

The relationship between efficiency and robustness proves to be structurally antagonistic. Because standby or redundant railway infrastructure is not operationally viable at scale, the edges that carry the network most efficiently are, by the same token, the ones whose removal causes the greatest disruption. A highly efficient connection — one through which a disproportionate share of shortest paths passes — is also the network's most fragile point, unless an equally efficient alternative route exists in its absence. There is no configuration that simultaneously maximises both properties; the practical question is always one of where on the efficiency–robustness frontier a given operational posture sits, and whether the trade-off is made deliberately or by default.

References

  1. Serrano, M. Á., Boguñá, M., & Vespignani, A.(2009).Extracting the multiscale backbone of complex weighted networks.Proceedings of the National Academy of Sciences, 106(16), 6483–6488.
  2. Radicchi, F., Ramasco, J. J., & Fortunato, S.(2011).Information filtering in complex weighted networks.Physical Review E, 83(4), 046101.
  3. Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E.(2008).Fast unfolding of communities in large networks.Journal of Statistical Mechanics: Theory and Experiment, 2008(10), P10008.
  4. Dijkstra, E. W.(1959).A note on two problems in connexion with graphs.Numerische Mathematik, 1(1), 269–271.
  5. Spearman, C.(1904).The proof and measurement of association between two things.The American Journal of Psychology, 15(1), 72–101.
  6. Deng, C., & Lin, H.(2014).Progressive and iterative approximation for least squares B-spline curve and surface fitting.Computer-Aided Design, 47, 32–44.
  7. Transport for London.(2024).NUMBAT — Network Usage by Mode, By Area and Time.TfL Open Data. crowding.data.tfl.gov.uk
  8. Transport for London.(2024).TfL Open Data platform.tfl.gov.uk/info-for/open-data-users/our-open-data
  9. OpenStreetMap contributors.(2024).OpenStreetMap.openstreetmap.org
  10. Chopra, S. S., Dillon, T., Bilec, M. M., & Khanna, V.(2016).A network-based framework for assessing infrastructure resilience: a case study of the London metro system.Journal of the Royal Society Interface, 13(118), 20160113.
  11. D'Lima, M., & Medda, F.(2015).A new measure of resilience: An application to the London Underground.Transportation Research Part A, 81, 35–46.
  12. Zhang, Y., Marshall, S., & Manley, E.(2019).Network criticality and the node-place-design model: Classifying metro station areas in Greater London.Journal of Transport Geography, 79, 102485.
  13. Strano, E., Shai, S., Dobson, S., & Barthelemy, M.(2015).Multiplex networks in metropolitan areas: generic features and local effects.Journal of the Royal Society Interface, 12(111), 20150651.
  14. Munoz-Mendez, F., Han, K., Klemmer, K., & Jarvis, S.(2018).Community Structures, Interactions and Dynamics in London's Bicycle Sharing Network.UbiComp/ISWC '18 Adjunct, 1015–1023.
  15. Barrat, A., Barthélemy, M., Pastor-Satorras, R., & Vespignani, A.(2004).The architecture of complex weighted networks.Proceedings of the National Academy of Sciences, 101(11), 3747–3752.
  16. Zhang, X., Zhang, Z., Zhao, H., Wang, Q., & Zhu, J.(2014).Extracting the Globally and Locally Adaptive Backbone of Complex Networks.PLoS ONE, 9(6), e100428.
  17. Zhao, J., Rahbee, A., & Wilson, N. H. M.(2007).Estimating a Rail Passenger Trip Origin-Destination Matrix Using Automatic Data Collection Systems.Computer-Aided Civil and Infrastructure Engineering, 22(4), 376–387.
  18. Beltjens, P.(2020).Passenger flows, backbone and robustness in subway networks.Master's Thesis, École polytechnique de Louvain, Université catholique de Louvain.
  19. Taboada, G. L., & Han, L.(2020).Exploratory Data Analysis and Data Envelopment Analysis of Urban Rail Transit.Electronics, 9(8), 1270.
  20. Kung, K. S., Greco, K., Sobolevsky, S., & Ratti, C.(2014).Exploring Universal Patterns in Human Home-Work Commuting from Mobile Phone Data.PLoS ONE, 9(6), e96180.
  21. Forero-Ortiz, E., & Martínez-Gomariz, E.(2020).Hazards threatening underground transport systems.Natural Hazards, 100(3), 1243–1261.
  22. Stone, M., & Aravopoulou, E.(2018).Improving journeys by opening data: the case of Transport for London (TfL).The Bottom Line, 31(1), 2–15.
  23. Song, C., Havlin, S., & Makse, H. A.(2005).Self-similarity of complex networks.Nature, 433, 392–395.