Source-linked AI summary
Towards a comparative science of cities: using mobile traffic records in New York, London and Hong Kong
S. Grauwin, S. Sobolevsky, S. Moritz, I. Gódor, C. Ratti
TL;DR
Earlier land-use studies often focused on one city and one kind of digital data, leaving questions about cross-city and cross-activity patterns. This chapter analyzes multiple mobile traffic counters across New York, London, and Hong Kong, finding shared activity signatures in their core business centers but distinct residential signatures.
Problem
Many land-use studies were limited to one spatial entity and one type of digital data, leaving open how mobile activity patterns compare across data types and cities.
Source: §1 · p. 2Method
The chapter analyzes time- and space-resolved mobile counter data from New York, London, and Hong Kong and uses clustering to identify locations with similar activity signatures.
Source: §1 · p. 2, §6 · p. 20Results
The three cities' core business centers fall into one cluster, while their residential locations have distinct signatures.
Source: §6 · p. 20Takeaways & Limitations
The authors suggest that globalization shapes activity in financial and decisional cores, while local factors remain prominent in individual activity patterns.
Source: §6 · p. 20Interpreting automatically generated insights from mobile traffic data requires knowledge expertise.
Source: §6 · p. 20
Abstract
from arXiv · showhide
This chapter examines the possibility to analyze and compare human activities in an urban environment based on the detection of mobile phone usage patterns. Thanks to an unprecedented collection of counter data recording the number of calls, SMS, and data transfers resolved both in time and space, we confirm the connection between temporal activity profile and land usage in three global cities: New York, London and Hong Kong. By comparing whole cities typical patterns, we provide insights on how cultural, technological and economical factors shape human dynamics. At a more local scale, we use clustering analysis to identify locations with similar patterns within a city. Our research reveals a universal structure of cities, with core financial centers all sharing similar activity patterns and commercial or residential areas with more city-specific patterns. These findings hint that as the economy becomes more global, common patterns emerge in business areas of different cities across the globe, while the impact of local conditions still remains recognizable on the level of routine people activity.
1 Introduction
Prior research used telecommunications traces to study individual mobility and to characterize urban landscapes and land use. This chapter compares multiple types of mobile activity across three cities, addressing the limited cross-city, cross-data comparisons in earlier land-use studies.
Prior work
- Earlier studies used mobile traces to identify regular individual routines and statistical regularities in human trajectories.This work contributed to a broader research field on human mobility from digital traces.Source: §1 · pp. 1–2
- Location-aggregated mobile records have been used to characterize urban landscapes, delineate regions, estimate population density, and identify social groups and events.One study clustered Rome grid pixels by usage patterns and qualitatively linked them to local business counts.Source: §1 · p. 2
Research gap
- Many land-use classification studies examined only one spatial entity and one type of digital data.Prior work used mobile calls or SMS, taxi trips, or Twitter data with methods ranging from clustering to neural networks.Source: §1 · p. 2
Research questions and approach
- The chapter asks whether patterns differ across calls, SMS, and data transfers, and how mobile network usage compares across New York, London, and Hong Kong.It analyzes a multi-modal collection of counter data resolved in time and space, then examines city-level dynamics and clusters local activity patterns.Source: §1 · p. 2
2 Materials and Methods
The study compares three globally connected cities with different urban geographies using aggregated mobile traffic and common 500 m grid cells. The data cover three months in 2013, with coverage gaps and processing choices that shape the spatial analysis.
2.1 Geographical background
- London, New York, and Hong Kong have distinct administrative and urban geographies, yet are comparable in size and population density and are all classified as Alpha cities.London's dataset is concentrated in Inner London, while Hong Kong's measurements cover urban zones rather than unpopulated mountainous areas.Source: §2.1 · pp. 2–3, Figure 1 · p. 3, §2.2 · p. 4
2.2 Data gathering and pre-processing
- Aggregated 3G traffic from several operators covered April 1 to July 7, 2013, at 15-minute intervals, across several million subscribers.Coverage spans all of New York, is concentrated in Inner London, and includes Hong Kong's urban zones.Source: §2.2 · p. 4
- The cell-level counters record calls, SMS, data requests, and uploaded and downloaded bytes, without revealing individual user information.Providers aggregated the data and applied a secret scaling factor, leaving only normalized amounts for each counter type.Source: §2.2 · p. 4
- The researchers harmonized differing city land-use categories into seven types and projected mobile, census, and land-use data onto 500 m by 500 m grid cells.They smoothed cell activity using the mean of cells within a 1,500 m square and selected the grid size to balance noise reduction and spatial detail.Source: §2.2 · pp. 4–5
3 Spatial repartition of activity
Mobile request activity is spatially concentrated in all three cities, with different center–periphery arrangements. Its inequality varies more by city than by activity type, and the analysis notes a coverage mismatch for London.
Spatial distribution
- Request activity shows a center–periphery dichotomy: Greater London is concentric, New York polycentric, and Hong Kong has a large center spanning Kowloon and northern Hong Kong Island, plus secondary centers in newly developed New Territories zones.Figure 2 maps normalized total request activity aggregated over three months on pixel grids.Source: §3 · p. 5, Figure 2 · p. 6
Spatial concentration
- Hong Kong has less than 2% of request activity in its 50% least-active pixels and 64% in its top 10%; London has around 9% and 49%, and New York around 9% and 38%, respectively.Figure 2(d) shows Lorenz curves of cumulative request activity against cumulative area, with pixels ranked by activity.Source: §3 · p. 6, Figure 2 · p. 6
- Gini coefficients vary more strongly by city than by activity type, ranking spatial inhomogeneity highest in Hong Kong, then London, then New York.The coefficients summarize the inequality of spatial activity distributions.Source: Table 2 · p. 5, §3 · p. 6
Limitations
- The analysis is limited by a mismatch between the area covered by the London dataset and Greater London's official boundaries.The authors nevertheless report that the initial analysis provides insights into how people interact with their cities.Source: §3 · p. 7
4 Exploring temporal patterns
Typical-week mobile signatures reveal a shared daily rhythm across New York, London, and Hong Kong, alongside local patterns that correspond to place types in London. The chapter describes city-signature differences as reflecting technological, economic, and cultural influences.
4.1 Typical week signature
- Typical-week profiles average each 15-minute interval across matching weekdays over three months, exclude civic holidays, and normalize pixel activity by its weekly mean.Normalization supports comparisons of relative temporal patterns despite different activity amplitudes.Source: §4.1 · p. 7
- Typical-week signatures capture main temporal patterns while reducing noise and long-term trends; residuals are typically 5–10% of timelines.Figure 3 decomposes Greater London’s mean-normalized Calls timeline into a repeating week pattern and residuals.Source: §4.1 · p. 8, Figure 3 · p. 8
4.2 Comparing Cities’ Signatures
- Across all three cities and activity components, mobile activity rises from 6 to 10 AM, stays fairly steady by day, and declines more slowly from 9 PM to 2 AM.The pattern is weaker on weekends, with smaller workday peaks around commuting and lunch times.Source: §4.2 · p. 8, Figure 4 · p. 9
- London’s request signature varies more over the day and drops more on weekends than New York’s or Hong Kong’s; evening data drops may reflect home Wi-Fi use.The authors suggest London’s higher mobile-data-plan costs may encourage switching to Wi-Fi, which the dataset does not record.Source: §4.2 · pp. 8–9
- New York and Hong Kong show evening-skewed activity, but Hong Kong SMS falls earlier while New York has evening SMS peaks; New York and Hong Kong appear more similar than New York and London.The authors caution that signatures reflect technological, economic, and cultural influences.Source: §4.2 · p. 10
4.3 Comparing Local Signatures
- City of London signatures fit a weekday business-area pattern, while Piccadilly remains active on weekends and ramps up smoothly in the morning, matching its commercial and tourist character.The City’s activity is very low on weekday evenings and has large weekday-to-weekend ratios.Source: §4.3 · pp. 10–11
- Camden shows weekday-evening SMS peaks and weekend lunchtime request activity; Newham’s residential pattern also appears in Ealing, which adds commute-hour request peaks.Ealing’s commuting-hour peaks align with its role as a travel hub.Source: §4.3 · pp. 11–12
- Local signatures correspond to the nature of places: London’s financial, commercial, recreational, residential, and commuting areas show distinct activity patterns.The five 500m by 500m pixels display a comparable day–night cycle alongside these location-specific characteristics.Source: §4.3 · pp. 10, 12, Figure 5 · p. 11
5 Cluster Analysis
K-means clustering of mobile-traffic signatures reveals cohesive urban areas with interpretable activity profiles. Across the three cities, core business centers share a pattern, while residential clusters retain city-specific differences.
5.1 Principles
- K-means groups pixels by similarity in five activity signatures, representing each pixel with a 3,360-dimensional vector across 672 weekly intervals.The distance measure does not account for the temporal order of intervals.Source: §5.1 · pp. 12–13
- The researchers selected K = 6 using silhouette-index local maxima; larger K mostly added small clusters in low-activity areas without regular signatures.Choosing the best K is difficult because it depends on data shape, scale, and desired clustering resolution, with no consensus on a best method.Source: §5.1 · pp. 13–15
5.2 Revealing the spatial structures of cities
- Within each city, six clusters form spatially cohesive, roughly concentric structures, including business cores, commercial areas, residential areas, and mixed areas.The business-core and residential interpretations align with census land-use data; other clusters vary by city.Source: §5.2 · pp. 16–18
- Other cluster interpretations differ by city: London’s pink cluster resembles residential areas but peaks at commuting hours, New York’s corresponds mostly to JFK, and Hong Kong’s is mixed residential.London’s pastel cluster is mainly residential; New York’s corresponds to parks with low activity shares and higher weekend activity; Hong Kong’s appears residential.Source: §5.2 · p. 18
5.3 Revealing Universal Patterns
- Core business centers from all three cities join one cluster with a shared activity signature, while residential clusters largely remain city-specific.Residential patterns reflect local cultural, technological, and economic factors.Source: §5.3 · p. 19
6 Discussion
London, New York, and Hong Kong share a clustered core-business signature, while residential signatures remain distinct. The discussion also presents mobile-traffic maps and typical-week signatures as complements to conventional land-use information.
6 Discussion
- Time-aggregated traffic maps locate where people are most active and complement census records of where they live or work; the authors describe them as more current and cheaper to obtain.They expect activity to correlate more with job density during working hours and with residential density when people are typically at home.Source: §6 · pp. 19–20
- Typical-week signatures reflect cultural, technological, economic, and mobile-plan factors across the three cities and can distinguish local area types.The authors suggest operators could use typical usage to inform network decisions and retailers could use vicinity patterns to schedule activity.Source: §6 · p. 20
- Clustering identifies business, residential, mixed-use, and other specialized areas, offering a relatively quick, inexpensive way to complement land-use maps with observed activity.Examples include commuting hubs in London and JFK airport in New York.Source: §6 · p. 20
- The three cities’ core business centers fall into one cluster, while their residential locations have distinct signatures, consistent with globalized cores and locally shaped activity.The authors frame this interpretation as tentative and leave the universality of the findings open.Source: §6 · p. 20
- Mobile-traffic insights still require expertise to understand, interpret, and critique rather than relying on automatic procedures alone.Source: §6 · p. 20
- Whether the reported similarities generalize beyond these three cities remains unresolved; the authors propose including cities in both developed and developing countries.Source: §6 · pp. 20–21