Content uploaded by Thiago H. Silva
Author content
All content in this area was uploaded by Thiago H. Silva on Mar 19, 2025
Content may be subject to copyright.
Journal of Internet Services and Applications, 2025, 16:1, doi: 10.5753/jisa.2025.5152
This work is licensed under a Creative Commons Attribution 4.0 International License.
Modeling Interest Networks in Urban Areas: A Comparative
Study of Google Places and Foursquare Across Countries
Gustavo H. Santos[Univ. Tecnológica Federal do Paraná |gustavohenriquesan-
tos@alunos.utfpr.edu.br]
Fernanda R. Gubert[Univ. Tecnológica Federal do Paraná |fernandagubert@alunos.utfpr.edu.br]
Myriam Delgado [Universidade Tecnológica Federal do Paraná |myriamdelg@utfpr.edu.br ]
Thiago H. Silva [Universidade Tecnológica Federal do Paraná |thiagoh@utfpr.edu.br ]
DAINF, Universidade Tecnológica Federal do Paraná (UTFPR), Av. Sete de Setembro, 3165, Rebouças, Curitiba,
PR, 80230-901, Brazil
Received: 03 November 2024 •Accepted: 25 February 2025 •Published: 17 March 2025
Abstract Location-Based Social Networks (LBSNs) are valuable for understanding urban behavior and providing
useful data on user preferences. Modeling their data into graphs like interest networks (iNETs) offers important
insights for urban area recommendations, mobility forecasting, and public policy development. This study uses
check-ins and venue reviews to compare the iNETs resulting from two distinct LBSNs, Foursquare and Google
Places. Although these two LBSNs differ in nature, with data varying in regularity and purpose, their resulting
iNETs reveal similar urban behavior patterns. When analyzing the impact of socioeconomic, political, and geo-
graphic factors on iNET edges — each edge representing users’ interests in a pair of regions — only geographic
factors showed a significant influence. When studying the granularity of area sizes to model iNETs, we highlight
important trade-offs between larger and smaller sizes. Additionally, we propose a methodology to identify clusters
of geographically neighboring areas where user interest is strongest, which can be advantageous for understanding
urban space usage.
Keywords: Location-Based Social Networks, Google Places, Foursquare, User Interest, Urban Areas
1 Introduction
Location-Based Social Networks (LBSNs) help us under-
stand several issues in the context of urban computing [Fer-
reira et al., 2015; Santala et al., 2017; Silva et al., 2017a;
Ladeira et al., 2019; Veiga et al., 2019; Ferreira et al., 2020;
Senefonte et al., 2022; Silver and Silva, 2023; Silva and Sil-
ver, 2024]. In particular, LBSNs offer urban data that inher-
ently reflect social aspects, such as user preferences [Silva
et al., 2019].
Geolocated user activities on LBSNs provide useful urban
data, such as reviews and check-ins at venues throughout
the city being analyzed. These activities can be aggregated
into an undirected graph in which nodes represent areas (e.g.,
neighborhoods) in the city where the data have been shared,
and each edge connects a pair of areas visited by the same
user. This type of modeling leads to what we refer to in this
work as Interest Networks (iNETs).
Modeling data through iNETs could provide valuable in-
sights to improve the understanding of user behavior in ur-
ban environments. For instance, it could provide a deeper
understanding of users’ interest in specific physical spaces
within the city, facilitate geographic area recommendations,
enhance mobility forecasting, and support the development
of public policies to increase interest in certain urban areas.
In this context, the main objective of this work is to
compare iNETs modeled using data from two LBSNs:
Foursquare and Google Places. From Foursquare, we uti-
lize check-in data, representing users sharing their location
at a venue with friends. In Google Places, we use data from
user reviews of venues. Both platforms capture the type of
venue the user is at, for example, a restaurant or a bookstore.
However, the data they provide are different: check-ins are
more real-time, reflecting where the user is at a specific mo-
ment, whereas reviews may be posted after the user has left
the venue. With this, we aim to investigate whether differ-
ent LBSNs capture similar information when modeling data
through iNETs, considering various variables and scenarios.
In our previous work [Santos et al., 2024], we compare
the two LBSNs in the context of Curitiba, Brazil, to deter-
mine whether they model urban behavior similarly in captur-
ing users’ interest in neighborhoods. The present work sig-
nificantly expands the previous similarity analysis between
LBSNs by considering areas in different countries. Addition-
ally, it assesses the impact on results and practical applica-
tions when varying the granularity of urban areas. Besides,
it shows how the iNETs can be used to better understand ur-
ban phenomena, examining the influence of socioeconomic,
political, and geographic factors on the interest for the neigh-
borhoods in Curitiba and aggregates analyses regarding Ur-
ban Preference Zones (UPZones), and their corresponding
networks (UPZ-iNETs) in London. These case studies also
show how iNETs can be used for urban analysis across coun-
tries.
The main contributions of this work can be summarized
as follows.
• We propose a new approach, Interest Networks (iNETs),
Modeling Interest Networks in Urban Areas Santos et al., 2025
to studying urban phenomena by analyzing user activi-
ties on Location-Based Social Networks (LBSNs). iN-
ETs offer innovative ways to explore how people en-
gage with different areas of a city.
• An impact analysis of urban area granularity on iNETs:
In the present work, we introduce the tool h3-cities1,
which enables the division of any city, with available
OpenStreetMap data, into hexagons of varying sizes.
Our findings show that when modeling iNETs at larger
granularities, the iNETs derived from the two LBSNs
are more similar, whereas smaller granularities reveal
greater differences. However, we should not only aim
for large granularity because we could miss important
nuances regarding user behavior – we present trade-offs
to guide further studies.
• An influence analysis of socioeconomic, political, and
geographic factors on iNETs: We observe that average
monthly income, racial composition, and political po-
larization do not significantly impact people’s interests
in urban areas. However, geographic distance moder-
ately correlates with urban behavior, as people often
visit nearby areas.
• A methodology for identifying urban preference zones
(UPZones) and their resulting networks: This method-
ology explores iNETs to identify clusters of geographi-
cally neighboring areas where user interest is strongest.
The results are consistent across LBSNs, even at smaller
granularities. This new methodology can be valu-
able for providing UPZone networks, referred to UPZ-
iNETs, and for future analyses in understanding urban
user interests.
The remainder of the article is organized as follows. Sec-
tion 2 reviews related work, while Section 3 describes the
used datasets and their characteristics. Section 4 details the
methods applied and developed for analyzing iNETs. Sec-
tion 5 presents the results. Finally, Section 6 concludes the
study and presents directions for future work.
2 Related Works
This section reviews relevant literature across four key top-
ics: comparison of LBSN data, people’s movement patterns,
understanding an urban context through different granulari-
ties, and the identification of similar urban zones.
2.1 LBSN Data Comparison
This section presents some works that use and compare data
from LBSNs in their research. For example, Silva et al.
[2013] investigated the possibility of using two different LB-
SNs, Instagram and Foursquare, to collect data for which lo-
cation was shared. The study sought to understand whether
the information obtained could be complementary and/or
similar for both bases regarding city dynamics and urban
behavior patterns. From this, they concluded that the two
datasets provided compatible and complementary informa-
tion, in which, for example, a check-in from Foursquare
1https://h3-cities.streamlit.app/
could bring information about the category of a venue com-
mented on in an Instagram post and also capture urban as-
pects in a similar way, such as the most popular areas of
cities.
Martí et al. [2019] explore the potential of using data from
LBSNs such as Foursquare, Twitter, Google Places, Insta-
gram, and Airbnb for research into urban phenomena, recog-
nizing not only the benefits but also the challenges that the
use of these data sources entails. The study presents research
that uses data from LBSNs to analyze the city’s dynamics and
proposes a methodology for data retrieval, selection, classi-
fication, and analysis. It also identifies the main thematic
lines of investigation based on the data provided by these
platforms to offer a framework for studying urban phenom-
ena through LBSN data.
Nolasco-Cirugeda and García-Mayor [2022] show how
data from Foursquare, Twitter, and Google Places are cru-
cial for analyzing the use of urban space and social dynam-
ics. By detailing pioneering research and case studies over
the last decade, the article shows how LBSN data offer impor-
tant information about urban life, helping to understand so-
cial dynamics and specific urban interventions. The research
employs multiple analytical approaches at scale to address
diverse urban issues, from social neighborhood dynamism
to tourism and green infrastructure preferences, highlighting
the comprehensive analytical potential of LBSN data.
Skora et al. [2022] examined whether information ex-
tracted from Foursquare data could resemble information
released by the WTO (World Tourism Organization). The
study found the potential of using LBSNs to facilitate the
understanding of tourist movements on larger scales and in
more detail than traditional sources, despite limitations asso-
ciated with LBSNs, such as the predominant use of young
people with internet access.
Our work does not seek to investigate the complementary
character of LBSNs, which could be beneficial for, for ex-
ample, data integration [Silva and Fox, 2024] nor compare it
with official sources; instead, it aims to seek to what extent
distinct LBSNs will model urban behavior in a similar way,
understanding when to expect different results when using
these tools.
2.2 Understanding Urban Displacement
Other studies use large-scale data, including LBSN data, and
data mining techniques to understand which factors may be
associated with people’s movement patterns. For example,
Cheng et al. [2021] used geolocated data from Twitter to un-
derstand user movements. The authors associated this spa-
tial information with the economic characteristics of users,
the geographic aspects of the areas frequented, as well as
their positioning within the social network and the language
used in their check-ins. Thus, they identified the Lévy Flight
model in mobility patterns, in which short distances are trav-
eled more frequently, and longer distances occur more rarely.
They also presented the influence of population density and
popularity on the social network on the distances traveled by
users.
Huang and Butts [2023] investigated several socioeco-
nomic characteristics and sought to understand which im-
Modeling Interest Networks in Urban Areas Santos et al., 2025
pacted migration between counties in the United States. This
work examined the hypothesis that migrations occurred be-
tween similar areas, proposing a theory of segregation and
social immobility about these movements. For their analy-
ses, the authors use a temporal graph model. In this type of
model, specific parameter settings allow an analysis of peo-
ple’s behavior in simulations where, for example, political
segregation is disregarded.
Santin et al. [2020] analyze public transit mobility patterns
of different economic classes in Curitiba, Brazil, using smart
card data. The authors find that higher-income classes delay
morning travel by about two hours and have more localized
trips than lower-income groups. A transit mobility network
reveals distinct spatial and temporal patterns across classes.
The approach is validated by comparing it with household
travel surveys, offering a cost-effective method for urban mo-
bility studies and providing insights into the socioeconomic
factors influencing urban transit usage. These findings are
key to urban planning and sustainable development.
Senefonte et al. [2022] focus on understanding and explor-
ing what drives international tourists’ mobility patterns us-
ing data from LBSNs. The authors construct mobility de-
scriptors, grouping users with similar behaviors based on
interests captured to places visits. The approach identifies
mobility patterns critical for tourism, revealing insights into
how tourists explore new countries based on their previous
travel profiles. The proposed approach enhances predictions
compared to traditional models, making it valuable for ur-
ban planners and tourism service providers in optimizing ser-
vices and improving tourist experiences through data-driven
insights.
Silva and Silver [2024] reveal how people’s movement pat-
terns between places such as restaurants, parks, and shops,
observed through Yelp, correlate with cultural attributes. The
authors model the interactions between locations and individ-
uals to study the cultural influences associated with urban
movement and social behavior. They also explore the po-
tential of using Graph Neural Networks (GNNs) to predict
cultural traits by analyzing Yelp data.
Our research aims to understand people’s interests in dif-
ferent city areas. Unlike these previous works, we do not cap-
ture and analyze users’ mobility patterns; instead, we map
users’ preferences across urban areas and seek to understand
these preferences through external factors.
2.3 Influence of Area Granularities
In this section, we discuss some of the most important pre-
vious studies that explore the need to understand the urban
context by analyzing different granularities. For example,
Rogov and Rozenblat [2018] noticed the absence of stud-
ies exploring cities’ resilience in their multi-scalar entirety.
While some works explored the ability to adapt to local im-
pacts (micro level), such as natural disasters, and their con-
sequences in the city (meso level), others investigated how
shocks in the city system (macro level) influenced the city
(meso level). They propose a framework for analyzing the
city across these three scales to ensure a comprehensive un-
derstanding. In this context, they discuss the role of interac-
tions between individuals (micro) that affect the urban char-
acteristics of the city (meso) and the impact on the overall
city system (macro). They also illustrate how, for example,
economic difficulties within this system (macro) influence
cities (meso) and alter individuals’ daily lives (micro). Thus,
they emphasize the importance of conducting studies at var-
ious levels of granularity.
Wu et al. [2020] employ the 2D discrete wavelet trans-
formation method to analyze the spatial structure of Beijing,
China, in a multi-scalar manner. To this end, they illustrate
how analyzing smaller areas captures the details of specific
locations but fails to describe the broader characteristics of
the city. They argue that analyzing different granularities
provides a holistic understanding of urban aspects, from area
definitions to their roles in the city’s overall structure.
Pafka [2022] shows how an urban analysis using statistical
and/or administrative divisions led to a series of studies that
are highly biased and difficult to compare. Therefore, they
use a grid system of squares of different sizes to analyze city
characteristics, such as the density of pedestrian-friendly ar-
eas and population density, in a standardized way in several
cities worldwide. In this way, they defend the importance of
consistently studying urban phenomena.
Based on those studies, we see the need to investigate the
urban context through different granularities. However, dif-
ferently from previous studies, we seek to understand the im-
portance of urban areas’ granularities in understanding urban
behavior in the context of interest per area.
2.4 Identification of Similar Urban Zones
Some studies have explored new methods for mapping ur-
ban zones by integrating multiple LBSN data sources to iden-
tify and classify urban spatial patterns and functional areas
[Cranshaw et al., 2012; Gao et al., 2017; Miao et al., 2021;
Shouji Du and Zheng, 2020; Ye et al., 2021].
For instance, Cranshaw et al. [2012] suggest a new way
of defining urban areas, the so-called Livehoods, dynamic
areas of activity, recognizing that the arbitrary delimitations
that form neighborhoods, for example, may not adequately
reflect the reality of a city to urban planning. Therefore, they
use data from Foursquare and a spectral clustering method
to define these new areas based on the venues’ similarity,
the proximity of check-ins and venues, and the behavior of
LBSN users. Furthermore, they validate these results in Pitts-
burgh through interviews with demographically diverse resi-
dents living in several parts of the city. They also used inter-
views with city hall professionals responsible for managing
public resources and professionals in the real estate market.
Gao et al. [2017] developed a method capable of extract-
ing topics from Foursquare data. These topics characterize
areas based on their functionalities and activities carried out
there. For example, a university and its surroundings can be
considered an educational region. They use a probabilistic
technique that classifies areas, considering the functions ex-
tracted directly from the data. As a result, the quantity and
specificity of topics/functionalities analyzed may vary.
By combining physical features from remote sensing im-
ages with social attributes from Point of Interest data, re-
searchers have achieved high accuracy in large-scale urban
zone mappings [Shouji Du and Zheng, 2020]. Additionally,
Modeling Interest Networks in Urban Areas Santos et al., 2025
integrating social media data and street-level imagery has
proven effective in recognizing urban functions, with verbs
extracted from social media posts serving as proxies for hu-
man activities [Ye et al., 2021]. These methods exemplify
the potential value of insights for urban planning, manage-
ment, and sustainability efforts regarding cities.
Unlike these studies, our proposal identifies urban zones
through a grid-based system, where user behavior determines
which cells are aggregated to form a zone, ensuring unifor-
mity in their construction.
3 iNET Datasources
This section explains how data can be obtained for modeling
iNETs, using information gathered from users located in dif-
ferent neighborhoods within the city of Curitiba, Brazil, Lon-
don, United Kingdom, as well as for 20 of the main American
counties and 20 of the main American cities. The section
also details the characteristics of the resulting datasets and
information about the collection of socioeconomic character-
istics of Curitiba neighborhoods. Curitiba has been chosen
because the authors are most familiar with the city, allowing
for a thorough investigation of the factors related to users’
interests. London was chosen because it is a city with a sig-
nificant amount of data in both LBSNs. The data from both
the LBSNs are more abundant in the USA, which is why they
were used across different cities.
3.1 LBSN Datasets
Google Places: The Google Places dataset is built by ex-
tracting reviews carried out by users of the Google Plus so-
cial network in venues registered with the Google Maps ser-
vice. The data are made available by the authors of He et al.
[2017] and Pasricha and McAuley [2018] for academic use.
From this dataset, we extract data from the city of Curitiba,
Brazil, London, United Kingdom, and American counties
(name, followed by state): New York, New York; Los An-
geles, California; Cook, Illinois; Clark, Nevada; Maricopa,
Arizona; District of Columbia, District of Columbia; San
Francisco, California; Harris, Texas; San Diego, Califor-
nia; Orange, Florida; Fulton, Georgia; Miami-Dade, Florida;
Philadelphia, Pennsylvania; Milwaukee, Wisconsin; Orange,
California; King, Washington; Kings, New York; Suffolk,
Massachusetts; Dallas, Texas; Travis, Texas; and Ameri-
can cities: New York, New York; Chicago, Illinois; Los
Angeles, California; Washington, District of Columbia; San
Francisco, California; Philadelphia, Pennsylvania; Houston,
Texas; Boston, Massachusetts; Atlanta, Georgia; Paradise,
Nevada; Austin, Texas; San Diego, California; Milwau-
kee, Wisconsin; San Antonio, Texas; Dallas, Texas; Seattle,
Washington; Indianapolis, Indiana; Phoenix, Arizona; Char-
lotte, North Carolina; Nashville, Tennessee. The Google
Places (G.P.) dataset includes specific information such as
the user’s name, education level, employment details, review
text (in multiple languages), review score, the time the re-
view has been posted, and a unique user identifier. This
dataset also contains specific information about the evaluated
venue, such as its name, category, opening hours, contact
phone number, address, latitude, and longitude.
Foursquare: We also use publicly available data, extracted
from Foursquare via check-ins shared on the social network
Twitter. Our Foursquare dataset is made up of user check-ins
in the cities of Curitiba, Brazil, and London, United King-
dom, as well as the aforementioned American regions. It
includes details such as the check-in date, the name and cat-
egory of the venue where it occurred, and the user’s unique
identifier. This dataset was explored and made available by
Silva et al. [2017b].
The amount and characteristics of available data for the
analyzed regions are presented in Table 1, for Google Places,
and Table 2 for Foursquare. These tables highlight that
Google Places has more reviews than Foursquare check-ins,
except in Curitiba, which shows fewer reviews than check-
ins. Additionally, Foursquare exhibits fewer distinct cat-
egories than Google Places. This is because Foursquare
uses more general categories (e.g., ”Food”), while Google
Places offers more specific ones (e.g., ”Italian Restaurant”).
Furthermore, Google Places data span a longer period than
Foursquare’s. It is important to note that although some cities
analyzed for Google Places contain data before 2010, the vol-
ume is significantly lower compared to the 2010–2014 pe-
riod. For this reason, we focus on the latter period to align
with the available Foursquare data.
Table 1. Description of Google Places dataset
Curitiba London USA
cities counties
Reviews 8,372 178,231 1,191,934 1,632,165
Users 4,909 75,897 394,588 486,393
Venues 2,213 31,075 186,639 286,075
Categories 685 1,81 3,439 3,793
Period from 2010 to 2014
Table 2. Description of Foursquare data
Curitiba London USA
cities counties
Reviews 53,253 27,088 398,805 495,698
Users 5,116 9,128 83,414 85,946
Venues 8,523 11,104 122,808 164,655
Categories 368 427 562 567
Period 2014* from 2012 to 2014
* from April to June
When comparing the most evaluated categories in Cu-
ritiba for both LBSNs, see Figure 1, the differences between
Google Places and Foursquare data are clear. For instance,
Foursquare includes information about users’ homes and
workplaces, which is absent in Google Places. This raises
an interesting question: can the modeled iNETs provide sim-
ilar insights despite these differences?
A comparison between the number of Google Places re-
views, and the number of Foursquare check-ins for each
county/city in the USA can be seen in Figure 2 for Ameri-
can counties and in Figure 3, for American cities. It is ev-
ident that Google Places contains a larger volume of data,
but Foursquare still provides a substantial amount of infor-
mation.
Modeling Interest Networks in Urban Areas Santos et al., 2025
(a) Google Places (b) Foursquare
Figure 1. Word Cloud of the 50 most reviewed or visited categories in
Curitiba
Since this work focuses on modeling through iNETs the
areas people frequent, we also examine how often users
post a message using their unique identifiers on each plat-
form. Additionally, we analyze the intervals between the
reviews/check-ins made by users. We then examine the num-
ber of users who made at least X publications (reviews/check-
ins) and the intervals between publications that occurred
within specific periods. These results are presented in Ta-
ble 3 for Google Places data and in Table 4 for Foursquare.
The tables illustrate user engagement with each platform and
present the intervals between reviews/check-ins in the ana-
lyzed areas.
It can be observed that Google Places users have a lower
frequency of platform usage than Foursquare users. Addi-
tionally, there is a noticeable difference in the regularity of
use between the two platforms. Google Places users tend to
post reviews either in quick succession or at intervals greater
than a week. In contrast, Foursquare users check in more
frequently and rarely go more than a week between check-
ins, particularly in Curitiba. Therefore, in addition to the dif-
ferences in the categories, we identify variations in the data
collected from the LBSNs concerning user engagement, par-
ticularly regarding the regular use of the tools provided by
each platform.
The previous analyses underscore the importance of this
study. Although the platforms Google Places and Foursquare
exhibit significant differences, if they model urban phenom-
ena similarly, they could reveal more general patterns about
the urban environment.
3.2 Socioeconomic and Electoral data
To understand the factors associated with people’s interests,
information on the socioeconomic aspects of each neighbor-
hood in Curitiba was obtained from the 2010 Brazilian Demo-
graphic Census conducted by the Brazilian Institute of Geog-
raphy and Statistics (IBGE)2. This national survey aimed to
portray the Brazilian population with its socioeconomic char-
acteristics and provide a basis for public and private planning
for the decade between 2000-2010. Aiming to understand
the differences between the neighborhoods of Curitiba, we
extract data on average monthly income and racial composi-
tion, the latter categorized as follows: White, Black, Brown,
Yellow, and Indigenous. This information has been collected
for all neighborhoods in Curitiba.
To gather electoral data in Curitiba, we utilize the informa-
tion provided by the Regional Electoral Court (TRE)3regard-
ing the second round of presidential elections in 2014. The
2https://sidra.ibge.gov.br
3https://www.tre-pr.jus.br/eleicoes/
eleicoes-anteriores/eleicoes- 2014
selection of these data is motivated by the works of Liu et al.
[2019] and Huang and Butts [2023], in which the discrep-
ancy in the percentages of votes during federal elections from
2004 to 2020, particularly in 2008 and 2020, was utilized to
assess political polarization between areas. For example, a
high percentage of votes for Democrats in one region and a
low percentage in another resulted in greater polarization.
The TRE data includes information on the number of vot-
ers who voted for each candidate at the polling place level,
organized into ten electoral zones. To extract data at the
neighborhood level, we tried a source from Curitiba city
hall that links voting locations to neighborhoods; however,
we only found this information for 2012 and 2016. Conse-
quently, to determine the percentage of voters who supported
a particular candidate in each neighborhood, we adopted the
Google Maps API to associate the addresses of voting loca-
tions (from a total of 418 voting locations) to their respective
neighborhoods, ensuring that the API results were specific
to the city of Curitiba. Through this method, we success-
fully linked voting locations to their corresponding neighbor-
hoods. Nonetheless, some neighborhoods lacked electoral
data, specifically: Centro Cívico, Campina do Siqueira, Alto
da Rua XV, Riviera, São Miguel, Caximba, Lamenha Pe-
quena, São João, and Cascatinha. These are shown hatched
in Figure 4c.
The socioeconomic data analyzed in the present work and
the electoral data collected for the experiments are shown in
Figure 4. The distribution of average income in the neighbor-
hoods of Curitiba is illustrated in Figure 4a. To represent the
racial composition, Figure 4b shows the percentage of people
self-declared Black for each neighborhood of Curitiba. Us-
ing the electoral data for the neighborhoods, we counted the
number of voters in each area and calculated the percentage
that voted for the government candidate and those that voted
for the opposition party. Political polarization can be visual-
ized in Figure 4c. It is possible to see in Figure 4c that some
of Curitiba’s neighborhoods are in peripheral areas, and, for
those who did not present data, their residents probably voted
in nearby neighborhoods that belong to the same electoral
zones. Although located in central areas, smaller neighbor-
hoods like Centro Cívico lack electoral information in this
study. One possible explanation is the same as that presented
for other areas - residents of these areas also voted in nearby
neighborhoods. Alternatively, the issue could stem from in-
accuracies in the method used to associate voting locations
with their respective neighborhoods.
In 2010, the income variation between the lowest and high-
est income neighborhoods was approximately sevenfold, and
the self-declared black population did not exceed 7.2% of the
total. In the 2014 elections, the percentage of voters for the
government candidate did not reach 45% in any of the neigh-
borhoods analyzed. Neighborhoods closer to the city center
exhibited a higher concentration of income and a predomi-
nantly white population, along with a lower percentage of
votes for the government candidate. Conversely, the highest
percentage of government votes occurred in the more remote
areas, particularly in the southern region, where neighbor-
hoods have lower purchasing power and a greater proportion
of self-declared Black residents.
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 2. Google Places reviews (Blue) vs. FourSquare check-ins (Orange) in American counties
Figure 3. Google Places reviews (Blue) vs. FourSquare check-ins (Orange) in American cities
Table 3. Description of User Interaction Patterns on Google Places
Recurrence in Reviews Interval Between Reviews
at least at least at least less than between between greater than
2 reviews 5 reviews 10 reviews 6 hours 6 and 24 hours 1 day and 1 week 1 week
Curitiba 23% 4.7% 1.49% 50.3% 2.93% 9.45% 37.3%
London 27.1% 7.8% 3% 56.9% 3.3% 9.42% 30.3%
USA cities 25.8% 7.18% 2.72% 57.9% 2.8% 7.53% 31.7%
USA counties 25.5% 6.8% 2.5% 55.1% 2.76% 7.71% 34.2%
Table 4. Description of User Interaction Patterns on Foursquare
Recurrence in Check-ins Interval Between Check-ins
at least at least at least less than between between greater than
2check-ins 5check-ins 10 check-ins 6 hours 6 and 24 hours 1 day and 1 week 1 week
Curitiba 76.9% 47.6% 28.1% 36.8% 25.5% 28.7% 9%
London 52.3% 15.7% 4.6% 55% 24.7% 20.3% 0%
USA cities 59.8% 23.5% 8.56 % 52.8% 27.8% 19.4% 0%
USA counties 61.2% 25.1% 9.4% 53.7% 27.7% 18.6% 0%
4 Methods
This section is organized into six topics. The first subsection
introduces iNETs. The second discusses methods for evalu-
ating the granularity of iNETs. The third subsection exam-
ines common patterns across different iNETs. The fourth ex-
plores the relationship between external influences and user
interests. The fifth subsection describes how these zones are
defined based on user preferences. Finally, the last subsec-
tion analyzes the common characteristics among these urban
zones.
4.1 Interest Networks (iNETs)
An iNET can help explain users’ interest in urban areas [San-
tos et al., 2024]. It is represented as a weighted, undirected
Modeling Interest Networks in Urban Areas Santos et al., 2025
(a) Average Monthly Income (R$) (b) Self-Identified Black People (%) (c) Votes for the Government Party Candidate (%)
Figure 4. Distribution of Socioeconomic and Electoral Characteristics in Curitiba
graph G= (V, E ), where the set Vof nodes represents the
city’s urban areas (such as grid cells, census tracts, zip codes,
neighborhoods, or boroughs). An edge ei,j ∈Econnects
urban area vi∈Vto urban area vj∈V, with the weight
wi,j ∈Nrepresenting the number of users who have re-
viewed venues in both areas. These edges, therefore, repre-
sent users’ interest in the two distinct urban areas. The con-
structed network also includes self-loops, meaning an edge
ei,i ∈Ethat connects a region vi∈Vto itself. The weight
wi,i ∈Non this edge represents the number of users who
have reviewed venues in the same area at least twice. As this
construction analyzes iNETs, it focuses only on edges with
positive weights. To construct iNETs, we use data from users
who reviewed at least twice, regardless of the venue
4.2 Testing iNETs Granularity
One of the objectives of this work is to investigate how
users’ interest in urban areas differs between the two LBSNs
(Google Places and Foursquare) and how the size and shape
of these areas affect this comparison. We developed a tool,
called h3-cities4, which utilizes OpenStreetMap and Uber’s
Hexagonal Hierarchical Geospatial Indexing System to sub-
divide a particular region into hexagons of various sizes.
This allows for the division of any city with available Open-
StreetMap data, providing a consistent, multi-scalar frame-
work for urban analysis, as emphasized by Rogov and Rozen-
blat [2018] and Pafka [2022] in their review of the litera-
ture. For our analyses in Curitiba, Londres and USA us-
ing h3-cities, we tested four different hexagonal grid reso-
lutions (hr): h6with an average area of 36.12km2,h7with
5.16km2,h8with 0.74km2, and h9with 0.11km2.
In addition to these areas delimitations, we consider some
4https://h3-cities.streamlit.app
extra boundaries in specific analysis. In USA regions, we
also use Census Tracts and Zip Codes. Figure 5 shows the
region of Chicago with the highest amount of data on Google
Places across all six levels of granularities used for that coun-
try. For Curitiba, we also used neighborhood delimitations.
A comparison between the neighborhoods and the hexago-
nal system at resolution h8is shown in Figure 6. We also
used London’s Boroughs for visualization purposes as will
be show in Figure 15.
Figure 5. Comparison of different granularity levels in Chicago
Census tract and zip code delimitations have been obtained
from the U.S. Census Bureau’s Tiger geographic database5;
and the neighborhood polygons from the Curitiba city hall
5https://www.census.gov/cgi-bin/geo/shapefiles/index.
php
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 6. Comparison of the segmentation of Curitiba through neighbor-
hoods vs. h8granularity level - a subdivision via h3-cities
website6. London’s boroughs were collected from the Lon-
don Datastore7.
4.3 Similarities Between the iNETs
To understand the similarity between the iNETs modeled
by the different LBSNs, we employed Pearson correlation,
which evaluates the presence and intensity of a linear rela-
tionship between two variables. Values close to 1 indicate
a strong positive linear relationship, whereas values close to
-1 indicate a strong negative linear relationship. When com-
paring the networks, we associate the weight of an edge in
the Google Places network with its respective edge in the
Foursquare network. Then, we calculate the Pearson correla-
tion using the entire set of edges in both graphs. We also use
the Spearman correlation for the edge weights, similar to the
Pearson correlation.
Furthermore, we verified the similarity to the eigenvector
centrality of both networks. This method indicates the im-
portance of a node in the network based on the importance
of the nodes with which it is connected. With it, we built
a ranking in the analyzed granularities, and to investigate
the similarity, we used Kendall’s Tau correlation coefficient.
This coefficient is used to evaluate the similarity between
the orders of two datasets. Values close to 1 indicate a sim-
ilar ordering, while values close to -1 suggest very different
orderings. With these three methods, we have different ways
to assess the similarity of the iNETs modeled by the differ-
ent LBSNs, getting a more robust analysis. This way, we can
check whether the interest (edges) are modeled similarly, as
well as whether the most important areas of the city (nodes)
are captured similarly.
4.4 Correlating External Factors and Interest
We also aim to understand whether additional factors—
particularly socioeconomic, political, and geographic affect
the interest of users from the analyzed areas in the two plat-
forms studied. For each region, we considered the following
6https://ippuc.org.br/geodownloads/geo.htm
7https://data.london.gov.uk/dataset/
statistical-gis- boundary-files- london
factors: average monthly income, racial composition, polit-
ical polarization, and geographic position. For brevity, we
show the results for Curitiba, but this method can be applied
to any city.
To achieve this goal, we correlate the edge weights with
the distances between a given factor. For average income,
we consider the absolute difference in average monthly in-
come between two neighborhoods. For political polarization,
we consider the results of the presidential election by region,
similar to the work of Huang and Butts [2023] and Liu et al.
[2019]. That is, we calculate the absolute difference between
the percentages that voted for the presidential candidate for
the areas. To compute the difference in racial composition,
we use the same method as Huang and Butts [2023], in which
the difference between neighborhoods Aand Bis defined as
RA,B =1
2
n
X
i=1
Pi(A)
P(A)−Pi(B)
P(B)
where RA,B is the difference in racial compositions between
neighborhoods Aand B,P(A)represents the population size
of neighborhood Aand Pi(A)the size of the population that
belongs to the i-th racial category in neighborhood A. We use
the categories defined by the 2010 Brazilian Demographic
Census: White, Black, Brown, Yellow, and Indigenous.
For the geographic distance, we use the centroids of each
neighborhood and calculate the distance using the Python li-
brary geopandas [Jordahl et al., 2020]. First, we transform
the geographic coordinates of latitude and longitude from the
WGS84 coordinate system to the UTM22S flat projection,
which provides the most accurate projection of the region
where Curitiba is located. Then, we calculate the geographic
distance between the two neighborhoods in meters, looking
at their centroids. The library uses the Euclidean distance for
the points in the flat projection and, based on the projection,
returns the geographic distance between the centroids of the
neighborhoods in meters.
4.5 Formation of Urban Preference Zones
With the iNETs of Google Places and Foursquare modeled,
we propose a new approach to defining urban preference
zones. When analyzing the strongest interests within the
iNETs—represented by the weight of the edges—we observe
that much of the interaction occurs between neighboring
areas. This suggests significant potential for using a grid
system to model zones according to users’ interests rather
than relying on arbitrary boundaries. Our proposal offers a
method for constructing zones of interest defined by the users
of the analyzed platforms, based on the places they frequent
and their spatial proximity.
To achieve this, we use iNET nodes to model densely con-
nected geographically adjacent zones. These zones consist of
cells with strong connections between them and spatial prox-
imity. Specifically, we select from the iNETs only the edges
between nodes whose hexagons share a vertex or edge, en-
suring that the captured interests reflect interactions between
nearby areas. In this work, we use the h9resolution to rep-
resent the cells (nodes in the graph) and compare the results
Modeling Interest Networks in Urban Areas Santos et al., 2025
obtained with this resolution against those using h8in Lon-
don. In this way, we can better understand the zones formed
in a detailed manner, but the approach is universal.
With this subgraph identified for each iNET, the Leiden
method [Traag et al., 2019] is applied to detect communities
within the graph, effectively capturing densely connected
nearby areas. This method ensures the identification of in-
ternally connected areas, and through multiple iterations, it
optimizes the communities found for maximum density. The
algorithm continues iterating on the graph until no further im-
provements are achieved. This process is illustrated in Fig-
ure 7. We also use the resolution parameter γ= 1, following
the value the method’s authors applied in their applications
[Traag et al., 2019]. Lowering this value would result in
larger areas while increasing it would produce smaller ones.
We use word clouds to improve the understanding of urban
preference zones. Similar to Hu et al. [2015] for semantic dif-
ferentiation of modeled areas, the natural language method
TF-IDF (Term Frequency - Inverse Document Frequency) is
used. This method emphasizes the importance of categories
that are frequently evaluated by people within an analyzed
area. At the same time, it reduces the relevance of categories
that are common in several areas. Thus, high TF-IDF val-
ues are given to categories that are more frequent in an area
but are also rarely mentioned in other areas. The goal is to
understand the differences between these areas by observing
the interests of users.
4.6 Similarity Between Urban Preference
Zones
We also aim to understand the similarities between the con-
structions of urban zones of interest among the studied LB-
SNs. After constructing the zones by clustering grid cells,
i.e., each cell is assigned to a community according to the
Leiden method, an alignment is performed between these
clusters using the Hungarian algorithm [Kuhn, 1955]. This
method seeks the optimal alignment of clusters to facili-
tate subsequent analyses, as cluster A in the Google Places
model may correspond to cluster B in Foursquare. Once this
alignment is completed, the Normalized Mutual Information
(NMI) metric is applied, where values close to 1 indicate sig-
nificant mutual information shared between the clusters, and
values near 0 indicate almost no mutual information. Addi-
tionally, the Rand Index [Hubert and Arabie, 1985] measures
the similarity between two clusters by evaluating all pairs of
samples and counting those that belong to the same clusters.
5 Results
This section is organized into four subsections. It covers
the iNETs modeled from Google Places and Foursquare data
and also the effect of different granularities on results in
cities worldwide. Then, it analyzes the influence of socioe-
conomic, political, and geographic factors on edge strengths,
showing results for Curitiba. Finally, it provides and dis-
cusses the identification of urban preference zones based
on user behaviors and their resulting networks (UPZ-iNET),
showing these results for London.
5.1 Modeled iNETs
To create the iNET for Google Places users in Curitiba, we
considered 1,127 users – recall that we consider users with
at least two reviews – and 4,590 reviews carried out by them.
We construct Foursquare’s iNET, considering 3,933 users
and their 52,033 check-ins. We present in Figure 8 the iN-
ETs generated for the city of Curitiba considering neighbor-
hoods to illustrate how the interest networks are formed. For
this case, we obtained a graph for Google Places comprising
69 neighborhoods (with available data) out of 75, containing
1,287 edges, 53 of which were self-loops. For Foursquare,
we modeled a graph with 75 nodes and 1,618 edges, 58 of
which are self-loops.
In the displayed graphs, the size of each node is propor-
tional to its weighted degree, i.e. the sum of the edge weights
connected to it, and the width of each edge is proportional
to its weight. The most noticeable visual difference is that
the network derived from Google Places shows interest in
areas slightly farther from the city center, such as the Alto
Boqueirão and Xaxim neighborhoods. Such a phenomenon
is not observed in the network modeled with Foursquare data.
Despite this, there is a considerable visual similarity between
the iNETs constructed from the two LBSNs. Both graphs re-
veal a high concentration of activities in the central region, in
neighborhoods such as Centro, Batel, Água Verde, Rebouças,
and Alto da Rua XV, where retail, restaurant, and office sec-
tors predominate. Other neighborhoods, such as Santa Feli-
cidade and São Francisco, are known for their numerous typi-
cal restaurants, bars, burger joints, and casual pubs with live
music. Meanwhile, neighborhoods like Bigorrilho, Centro
Cívico, Portão, Seminário, Jardim Botânico, Bacacheri, and
Cabral offer extensive green spaces with parks and squares,
as well as commercial and leisure infrastructure, including
shopping centers.
Regarding the other studied cities/counties, we modeled
the iNETs using the same methodology applied to Curitiba,
except for using other granularities instead of neighborhoods.
For London considering h9cells, for example, the Google
Places iNET comprises 123,292 reviews from 67,276 users,
forming 396,876 edges, including 1,533 self-loops, across
5,654 of the 6,600 cells covering the city. In contrast, the
Foursquare network comprises 22,738 check-ins from 8,178
users, forming a graph with 34,075 edges, including 1,209
self-loops, across 2,899 cells.
Regarding the United States, in total, we modeled 240 iN-
ETs for each LBSN – one for each of the 20 cities and 20
counties considered, at six different granularity levels. These
iNETs are very diverse. The iNET in the United States with
the lowest nodes is from New York county using h6cells
and Google Places data, with 7 nodes and the 28 edges be-
tween them, while the iNET with the most nodes is from Los
Angeles county using h9cells and Google Places data, with
10,363 nodes and the 455,976 edges between them.
5.2 Impact of Different Granularity Levels
To illustrate the differences between the studied LBSNs for
different granularity levels, Figure 9 shows the areas of
strongest user presence for each LBSN in the city of Chicago.
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 7. Formation of Urban Preference Zones (UPZones)
(a) The iNET from Google Places data (b) The iNET from Foursquare data
Figure 8. LBSN iNETs in the City of Curitiba
In this figure, darker colors indicate a higher number of
reviews/check-ins, whereas white represents the absence of
data. In addition to the spatial distribution of this informa-
tion, the figure also presents a cumulative distribution func-
tion (CDF) that provides a general overview of the amount
of data throughout the city. The figure illustrates the differ-
ences across several analyzed granularity levels, indicating a
considerable tendency for larger areas to be represented sim-
ilarly. Conversely, when smaller areas are used, there is a
tendency for greater variability among the LBSNs.
With the 240 iNETs, for each LBSN, modeled in the USA
(six granularities for each one of 20 cities and 20 counties),
we conducted similarity analyses between them, as explained
in Section 4.3 – one of them considers ranking. Before pre-
senting the comparison results, to better understand how the
ranking of urban areas is formed and compared, Figure 10
illustrates the process. It presents the 20 most central neigh-
borhoods in Curitiba, ranked by eigenvector centrality in
Google Places’ iNET, along with their respective rankings
in Foursquare’s iNET. Visually, small variations in the order
of the neighborhoods can be observed when comparing the
two datasets. However, the rankings remain relatively close,
indicating a moderate relationship between the centralities.
The results of the mentioned comparison are presented in
Figure 11 for three different correlation metrics: Pearson,
Spearmen, and Kendaltau. In this figure, the y-axis indicates
the correlation value analyzed for each granularity level,
while the x-axis presents, on a logarithmic scale, the number
of nodes in both iNETs (Google Places and Foursquare).
Observing Figure 11, there is an evident pattern that the
more nodes an iNET has, the greater the differences between
the iNETs modeled by the studied LBSNs. It is also impor-
tant to highlight that counties vary in area from 87km2to
23,895km2while cities vary from 177km2to 1,740km2.
As a result, some regions have more nodes at the h7granu-
larity level than other regions at the h8level, even if the h8
nodes are smaller than the h7.
Figure 12 shows the results of the similarity correlations
between iNETs from Google Places and Foursquare in Cu-
ritiba, with divisions into neighborhoods and h6to h9granu-
larity levels. It is worth noting that Curitiba’s neighborhoods
vary in size between h7and h8levels. The results for the iN-
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 9. The subdivisions from Google Places and Foursquare data in the
city of Chicago: analyses for six different granularity levels
Figure 10. Comparison of the most central neighborhoods of Curitiba con-
cerning centrality by eigenvector
ETs in London are shown in Figure 13 based on the analyzed
granularity levels. These findings reinforce the notion that
the smaller the urban areas analyzed, the greater the differ-
ences among the iNETs.
In the quest to understand users’ interests in different ur-
ban areas, these results provide crucial information. For in-
stance, when investigating the central areas of the city and
their relationships with other urban areas from a broader per-
spective, the granularity level h6, with areas of 36.12km2,
can be particularly important. This level tends to yield sim-
ilar results regardless of the chosen LBSN. In contrast, an-
alyzing the urban landscape at the h7and h8levels results
in smaller urban areas, which can provide more detailed in-
sights than larger areas obtained with h6. However, these
finer granularity levels may lead to the emergence of more
diverse iNETs, potentially resulting in a loss of relevant in-
formation and capturing only the behaviors of users specific
to each LBSN.
The h9granularity level exhibits the most significant dif-
ferences in the LBSNs. This occurs because h9areas are
much smaller, allowing for better differentiation between the
LBSNs. However, due to the limited data and a much larger
number of nodes in the formed iNETs, noise presence can be
more pronounced. When examining census tracts, zip codes,
as shown in figure 9, and neighborhoods, as illustrated in fig-
ure 8, we find that the census tracts and neighborhoods are
similar in the formed iNETs comparable to the h7and h8
granularity levels. In contrast, zip codes show a greater sim-
ilarity to the h6level in terms of homogeneity among iNETs
associated with both LBSNs. These findings align closely
with the results reported by Wu et al. [2020], as smaller ar-
eas capture different details compared to larger ones within
a city. Thus, each granularity offers distinct and complemen-
tary insights into urban phenomena.
An advantage of analyzing divisions used by government
entities is that their associated socioeconomic data help inves-
tigate factors related to users’ interest in urban areas. How-
ever, using different granularity levels provides, for example,
the flexibility to look at an urban landscape capturing overall
patterns, as in h6level, or examine the relationship between
small areas, with h9level.
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 11. Results of correlation between both iNETs in USA
5.3 Influence of Socioeconomic, Political and
Geographic Factors
This section aims to determine whether additional factors –
particularly socioeconomic, political, and geographic – af-
fect the interests of people who frequent the analyzed areas
Figure 12. Results of correlation between both iNETs in Curitiba
Figure 13. Results of correlation between both iNETs in London
through a study case in Curitiba. We exemplify this analysis
for Curitiba using neighborhoods as graph nodes due to their
association with socioeconomic data.
As explained in Section 4.3, we correlate the edge weights
of the modeled networks with the distances associated with
a given factor for each pair of nodes in the graph.
To gain further insight into how these distances apply to
the city of Curitiba, Figure 14 presents a subgraph show-
ing the distances for political and socioeconomic factors in
some of the most central neighborhoods in Curitiba. The
node color is proportional to the factor value being analyzed:
darker colors represent higher values, while lighter colors
correspond to lower values within the subset of the addressed
neighborhoods. For political polarization, darker nodes indi-
cate a higher percentage of votes for the government presi-
dential candidate. For racial composition, the percentage of
individuals identified as Black in each neighborhood is rep-
resented. For average monthly income, darker nodes corre-
spond to higher income levels. The edge color reflects the
difference between neighborhood factors: darker edges indi-
cate greater distances, while lighter ones represent smaller
differences. In Figures 14a, 14b and 14c, the edge differ-
ences account for the political, racial and income composi-
Modeling Interest Networks in Urban Areas Santos et al., 2025
tions, respectively, as previously explained.
Intending to gain deeper insight into the factors influenc-
ing visitors between pairs of nodes, we relate each edge
weight to its distance across the analyzed characteristics (av-
erage monthly income, racial composition, political polariza-
tion, geographic position), for both iNETs. Using Spearman
correlation, the results for both networks, considering infor-
mation from both complete iNETs (shown in Figure 8) and
filtered iNETs, are presented in Table 5, for Google Places
networks, and in Table 6, for Foursquare networks.
Table 5. Results of Spearman Correlations with the Analyzed Fac-
tors
Google Places Google Places
Complete Filtered
Average monthly income -0.09 -0.10
Racial Composition -0.25 -0.19
Political Polarization -0.23 -0.14
Geographic Distance -0.38 -0.30
Table 6. Results of Spearman Correlations with the Analyzed Fac-
tors
Foursquare Foursquare
Complete Filtered
Average monthly income 0.06 -0.04
Racial Composition -0.11 -0.20
Political Polarization -0.09 -0.15
Geographic Distance -0.55 -0.44
In complete iNETs, we observe the impact of all edges
found. In the filtered iNET, we only consider edges with a
weight greater than or equal to 5, focusing only on areas con-
nected by higher interest. For the filtered iNET, the Google
Places modeling shows only 418 edges out of the total 1,287
(32.5%) edges in the full network, while the Foursquare data
show 1,169 edges out of the total 1,618 (72.2%).
The results suggest that average monthly income, racial
composition, and political polarization do not significantly
explain visitors’ interest in specific areas. In other words,
people do not necessarily visit places with similar or dissimi-
lar income levels, racial demographics, or political views, as
reflected in the behavior captured by these LBSNs.
Across all scenarios, geographic distance shows the
strongest negative correlation, as seen in Tables 5 and 6. This
indicates that the greater the geographic distance between
two neighborhoods, the lower the edge weights. Notably, the
correlation with geographic distance is stronger in the iNET
modeled by Foursquare. A plausible explanation lies in the
differing ways users engage with LBSNs. This result aligns
with expectations, as several studies [Cheng et al., 2021;
González et al., 2008; Rhee et al., 2008; Brockmann et al.,
2006] highlight the tendency for users’ travel distances —
both in LBSNs and similar datasets — to cluster at shorter dis-
tances and become increasingly rare over greater distances,
as we have demonstrated.
The results for racial composition, though relatively weak,
show the second-highest correlation across all scenarios,
making it an interesting finding, as it was less expected com-
pared to geographic distance. Moreover, there is support-
ing evidence for this outcome. As shown in de la Prada
and Small [2024], racial differences between areas tend to
increase with distance, up to a threshold of 10 km. This in-
sight opens new possibilities for exploring other factors re-
sembling these factors and aims to explain interest in urban
areas.
For brevity, we focused the analysis in this section mostly
on the Curitiba results. However, our study indicates that
similar results concerning the influence of socioeconomic,
political, and geographic factors can be found in other cities
as well.
5.4 Urban Preference Zones (UPZones)
Using the method described in Section 4.5 applied in London,
with the h9 cells, we identified 1,760 UPZones on Google
Places and 1,023 on Foursquare. Figure 15 illustrates the
results of this process, compared with the divisions of the
city into boroughs. In this figure, each color represents a
distinct UPZone; however, identical colors in non-adjacent
areas indicate different zones. The UPZones are delineated
by borders, while the significantly larger areas, or boroughs,
are marked by gray boundaries. This comparison highlights
that constructing urban preference zones can provide more
detailed insights into user interest by better capturing the in-
tricacies of the urban landscape.
Figure 16 illustrates how the construction of UPZones
from the iNET with h9granularity level captures seman-
tically relevant areas within the city. In this figure, each
pair, color and number, corresponds to an UPZone based on
Google Places data, allowing us to observe their spatial dis-
tribution in London. The eight UPZones with the highest
activity levels have been selected (the numbers indicate their
ranking). Additionally, the figure features a word cloud that
visualizes the most prominent categories in each region (see
Section 4.6), highlighting categories that are more frequently
represented in each zone.
The figure illustrates that, for example, UPZone 1 encom-
passes one of the city’s most popular areas, known as Soho,
renowned for its vibrant nightlife and diverse array of restau-
rants. UPZone 2 is a celebrated region recognized for its
artistic activity, serving as a hub for artists and a venue for
various performances. Meanwhile, UPZone 3 represents the
area known as Mayfair, famous for its clothing and acces-
sories shops. UPZone 6 captures the South Bank, noted for
its cinemas, art galleries, and iconic tourist attractions like
the London Eye. The other UPZones also reflect well-known
areas visited by users with similar demographic interests. It’s
worth mentioning that the central area of the city was simi-
larly captured by Foursquare data, although the considerably
smaller dataset may affect this comparison. The proposed
method effectively identifies urban preference zones within
a city, enhancing our understanding of user preferences. By
combining this technique with insights about the areas and
conducting a detailed analysis, we can uncover regions in
the city that do not strictly adhere to predefined geographic
boundaries, thereby offering a richer understanding of urban
space utilization.
Modeling Interest Networks in Urban Areas Santos et al., 2025
(a) Political Polarization (b) Racial Composition
(c) Average Monthly Income
Figure 14. Sub-Graph of Distances Between Socioeconomic Factors
(a) Google Places
(b) Foursquare
Figure 15. Urban Preference Zones (UPZones) in the City of London, UK.
Different colors only indicate different zones. For comparison, official de-
limitations of boroughs are drawn in gray lines on the map
Figure 16. Characterization of the main UPZones found
The modeling of the city’s areas of interest was also ex-
plored using the cell resulting from h8granularity level as
Modeling Interest Networks in Urban Areas Santos et al., 2025
the iNET node, following the same method described earlier.
However, as shown in Figure 17, which compares the 8 zones
with the highest activity for each resolution (Red - h8; Brown
-h9), it becomes evident that, although both resolutions cap-
ture densely connected and nearby cells, the h8resolution
forms much larger zones. This results in a loss of detail and
subtle insights that the finer h9resolution provides.
Figure 17. Comparing resolution h8(red) and h9(brown) in modeling
UPZones
Using the metrics to compare UPZones defined in Section
4.6, we analyzed the urban scenarios constructed from the
Google Places and Foursquare iNETs. This analysis yielded
an NMI score of 0.6364. Additionally, the Rand Index ob-
tained from comparing the UPZones from Google Places and
Foursquare was 0.7378. These results indicate that when em-
ploying UPZones to model users’ interests, a notable simi-
larity is revealed despite the differences in the iNETs of the
studied LBSNs. Thus, it can be concluded that, regardless
of the LBSN utilized, moderately similar UPZones can be
derived.
For brevity, we presented results only for London. How-
ever, we also did the same experiment on Curitiba, and the
message observed is the same. In this way, we have an indica-
tion that UPZones can be applied to different cities, offering
an enhanced insight into urban space usage.
5.5 UPZ-iNETs: Interest Network of UP-
Zones
The urban preference zones also enable a qualitative analy-
sis of the interest between zones by visualizing which UP-
Zones of the city are most interconnected through user ac-
tivity. To achieve this, iNETs can be built with UPZones
as graph nodes, following the same methodology used to
construct the iNETs. This approach allows us to identify
zones that share mutual user interests. The resulting network,
referred to as UPZ-iNET, is shown in Figure 18 based on
Google Places data. To enhance readability and highlight
key connections, we retained only the top 0.1% of edges and
employed arrows for illustrative purposes. These undirected
edges represent the main links between zones, with weights
ranging from 203 to 1,456, corresponding to the number of
users who visited both areas. Thicker, darker edges denote
stronger connections, reflecting a higher number of shared
visitors between zones.
Figure 18 reveals a strong connection, for instance, be-
tween the central area and several other regions, including
non-central parts of the city. This visualization not only high-
lights the patterns of user interest between different areas but
also supports a qualitative analysis of the zones frequented by
the same users. Through this approach, we provide a method
for modeling urban areas in any city based solely on LBSN
user activity, providing insights into the interaction between
these urban spaces.
6 Conclusion
Understanding urban behavior is a complex task. In this
work, we explored whether two different Location-Based
Social Networks (LBSNs), Google Places and Foursquare,
could provide comparable insights when modeling users’
interests in geographic areas. We also examined how the
modeling of LBSN data influences the definition and under-
standing of urban areas. Through an analysis of the charac-
teristics of two datasets (resulting from Google Places and
Foursquare LBSN data) and information collected from Cu-
ritiba, London, and several U.S. cities and counties, we found
that the resulting graphs — referred to as Interest Networks
(iNETs) — effectively capture the dynamics of users’ behav-
ior. These iNETs exhibit significant similarities, particularly
in connections involving the most central urban areas at the
h6granularity level, while greater differences emerge when
smaller spatial units are employed for analysis (e.g., h8and
h9granularity levels).
Additionally, we investigated whether LBSN users’ inter-
est in urban areas could be understood through geographic
distance, political polarization, and socioeconomic charac-
teristics from the areas they visit. Our findings indicate
that, for the analyzed data, factors such as average income,
racial composition, and political polarization of the areas
(e.g., neighborhoods in Curitiba) do not sufficiently explain
users’ preferences. However, we observed that geographic
distance plays a limiting role in interactions, as users tend to
visit nearby areas.
Another aspect analyzed was the potential to minimize
the differences observed between the iNETs derived from
smaller areas (e.g., h8,h9). We proposed a method for defin-
ing urban preference zones (UPZones) within a city, empha-
sizing the capture of densely connected areas. The iNETs
formed from these zones, referred to as UPZ-iNETs, exhib-
ited greater similarity across the two LBSNs analyzed.
It is important to highlight that the use of LBSNs must be
accompanied by a critical understanding of the limitations
and implications of their applicability. For instance, data
from Google Places and Foursquare may not fully capture
the interest of the entire population, as users of these plat-
forms tend to be younger individuals with access to mobile
internet. Nonetheless, these platforms provide valuable in-
sights into the behavior of this demographic, which may or
may not reflect broader societal patterns. Additionally, it is
important to note that the data used in this study were col-
lected a decade ago, meaning that the results might differ if
more recent datasets were analyzed. However, the methodol-
ogy remains applicable, and since the data sources used are
publicly accessible, they offer a solid foundation for further
exploration in urban computing research.
Modeling Interest Networks in Urban Areas Santos et al., 2025
Figure 18. UPZ-iNET: Interest network of urban preference zones
In future research, it would be valuable to analyze more re-
cent datasets to determine whether different LBSNs continue
to yield similar insights and to explore how factors influenc-
ing iNETs may change over time. This exploration could
include examining additional variables such as cultural in-
fluences, types of venues, and the content of user reviews,
including sentiment analysis and topic modeling. These fac-
tors would enrich the analysis and help address questions
like whether the places frequented by users in urban envi-
ronments share cultural similarities, whether individuals are
inclined to visit the same types of venues across different
areas, and whether certain regions are characterized by spe-
cific venue categories. Furthermore, it would be interesting
to investigate whether the most interconnected areas within
an iNET reflect similar sentiments among users. This com-
prehensive approach would enhance our understanding of ur-
ban behavior and the dynamics of city life concerning venues
people commonly visit.
With a better understanding of the factors influencing the
interests of the populations in each city, these analyses could
be integrated into recommendation systems that highlight the
unique characteristics of each neighborhood. Additionally,
this type of investigation has the potential to inform public
bodies about the most interconnected areas, enabling the de-
velopment of public policies aimed at combating epidemics
or promoting social integration between previously discon-
nected regions. Furthermore, a deeper exploration of the
formation of urban preference zones, combined with valida-
tion from residents and experts in various cities, could en-
hance public policy systems for more effective management
of available resources.
Declaration
Acknowledgements
This research was partially supported by the SocialNet project
(process 2023/00148-0 of the São Paulo Research Foundation -
FAPESP), by the National Council for Scientific and Technological
Development - CNPq (processes 313122/2023-7, 314603/ 2023-9,
441444/2023-7, and 444724/2024-9). This research is also part of
the INCT of Intelligent Communications Networks and the Internet
of Things (ICoNIoT) funded by CNPq (proc. 405940/2022-0 ) and
Coordenação de Aperfeiçoamento de Pessoal de Nível Superior –
Brasil (CAPES) Finance Code 88887.954253/2024-00.
Funding
This manuscript did not receive any external funding.
Authors’ Contributions
GS performed the experiments. GS, FG, MD, and TS helped in the
conceptualization of the study and writing of the manuscript. GS is
the main contributor and writer of this manuscript. All authors read
and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Availability of data and materials
The tool h3-cities used in the study is available at: https://
h3-cities.streamlit.app/
References
Brockmann, D., Hufnagel, L., and Geisel, T. (2006). The
scaling laws of human travel. Nature, 439(7075):462–465.
DOI: 10.1038/nature04292.
Cheng, Z., Caverlee, J., Lee, K., and Sui, D. (2021). Ex-
ploring millions of footprints in location sharing ser-
vices. Proceedings of the International AAAI Con-
ference on Web and Social Media, 5(1):81–88. DOI:
10.1609/icwsm.v5i1.14109.
Cranshaw, J., Schwartz, R., Hong, J., and Sadeh, N. (2012).
The Livehoods Project: Utilizing Social Media to Under-
stand the Dynamics of a City. Proceedings of the In-
ternational AAAI Conference on Web and Social Media,
6(1):58–65. DOI: 10.1609/icwsm.v6i1.14278.
Modeling Interest Networks in Urban Areas Santos et al., 2025
de la Prada, A. G. and Small, M. L. (2024). How people
are exposed to neighborhoods racially different from their
own. Proceedings of the National Academy of Sciences,
121(28). DOI: 10.1073/pnas.2401661121.
Ferreira, A. P., Silva, T. H., and Loureiro, A. A. (2020). Un-
covering spatiotemporal and semantic aspects of tourists
mobility using social sensing. Computer Communications,
160:240–252. DOI: 10.1016/j.comcom.2020.06.005.
Ferreira, A. P. G., Silva, T. H., and Loureiro, A. A. F. (2015).
Beyond sights: Large scale study of tourists’ behavior us-
ing foursquare data. In 2015 IEEE International Confer-
ence on Data Mining Workshop (ICDMW), pages 1117–
1124. DOI: 10.1109/ICDMW.2015.234.
Gao, S., Janowicz, K., and Couclelis, H. (2017). Extracting
urban functional regions from points of interest and human
activities on location-based social networks. Transactions
in GIS, 21(3):446–467. DOI: 10.1111/tgis.12289.
González, M. C., Hidalgo, C. A., and Barabási, A.-L. (2008).
Understanding individual human mobility patterns. Na-
ture, 453(7196):779–782. DOI: 10.1038/nature06958.
He, R., Kang, W.-C., and McAuley, J. (2017). Translation-
based recommendation. In Proceedings of the Eleventh
ACM Conference on Recommender Systems, RecSys ’17.
ACM. DOI: 10.1145/3109859.3109882.
Hu, Y., Gao, S., Janowicz, K., Yu, B., Li, W., and
Prasad, S. (2015). Extracting and understanding ur-
ban areas of interest using geotagged photos. Comput-
ers, Environment and Urban Systems, 54:240–254. DOI:
10.1016/j.compenvurbsys.2015.09.001.
Huang, P. and Butts, C. T. (2023). Rooted america: Im-
mobility and segregation of the intercounty migration net-
work. American Sociological Review, 88(6):1031–1065.
DOI: 10.1177/00031224231212679.
Hubert, L. and Arabie, P. (1985). Comparing parti-
tions. Journal of Classification, 2(1):193–218. DOI:
10.1007/BF01908075.
Jordahl, K., den Bossche, J. V., Fleischmann, M., Wasser-
man, J., McBride, J., Gerard, J., Tratner, J., Perry, M.,
Badaracco, A. G., Farmer, C., Hjelle, G. A., Snow, A. D.,
Cochran, M., Gillies, S., Culbertson, L., Bartos, M., Eu-
bank, N., maxalbert, Bilogur, A., Rey, S., Ren, C., Arribas-
Bel, D., Wasser, L., Wolf, L. J., Journois, M., Wilson,
J., Greenhall, A., Holdgraf, C., Filipe, and Leblanc, F.
(2020). geopandas/geopandas: v0.8.1. DOI: 10.5281/zen-
odo.3946761.
Kuhn, H. W. (1955). The Hungarian method for the assign-
ment problem. Naval Research Logistics Quarterly, 2(1-
2):83–97. DOI: 10.1002/nav.3800020109.
Ladeira, L., Souza, A., Filho, G. R., Silva, T. H., and
Villas, L. (2019). Serviço de sugestão de rotas se-
guras para veículos. In Anais do XXXVII Simpósio
Brasileiro de Redes de Computadores e Sistemas Dis-
tribuídos, pages 608–621, Porto Alegre, RS, Brasil. SBC.
DOI: 10.5753/sbrc.2019.7390.
Liu, X., Andris, C., and Desmarais, B. A. (2019). Mi-
gration and political polarization in the u.s.: An analy-
sis of the county-level migration network. PLOS ONE,
14(11):e0225405. DOI: 10.1371/journal.pone.0225405.
Martí, P., Serrano-Estrada, L., and Nolasco-Cirugeda,
A. (2019). Social media data: Challenges, oppor-
tunities and limitations in urban studies. Comput-
ers, Environment and Urban Systems, 74:161–174. DOI:
10.1016/j.compenvurbsys.2018.11.001.
Miao, R., Wang, Y., and Li, S. (2021). Analyzing urban
spatial patterns and functional zones using sina weibo poi
data: A case study of beijing. Sustainability, 13(2). DOI:
10.3390/su13020647.
Nolasco-Cirugeda, A. and García-Mayor, C. (2022).
Social dynamics in cities: Analysis through lbsn
data. Procedia Computer Science, 207:877–886. DOI:
10.1016/j.procs.2022.09.143.
Pafka, E. (2022). Multi-scalar urban densities: from the
metropolitan to the street level. URBAN DESIGN Interna-
tional, 27(1):53–63. DOI: 10.1057/s41289-020-00112-y.
Pasricha, R. and McAuley, J. (2018). Translation-
based factorization machines for sequential recommen-
dation. In Proceedings of the 12th ACM Conference
on Recommender Systems, RecSys ’18. ACM. DOI:
10.1145/3240323.3240356.
Rhee, I., Shin, M., Hong, S., Lee, K., and Chong, S. (2008).
On the levy-walk nature of human mobility. In IEEE IN-
FOCOM 2008 - The 27th Conference on Computer Com-
munications. IEEE. DOI: 10.1109/infocom.2008.145.
Rogov, M. and Rozenblat, C. (2018). Urban Re-
silience Discourse Analysis: Towards a Multi-Level Ap-
proach to Cities. Sustainability, 10(12):4431. DOI:
10.3390/su10124431.
Santala, V., Miczevski, S., de Brito, S. A., Baldykowski,
A. L., Gadda, T., Kozievitch, N., and Silva, T. H. (2017).
Making sense of the city: Exploring the use of social me-
dia data for urban planning and place branding. In Anais
do I Workshop de Computação Urbana, Porto Alegre, RS,
Brasil. SBC. Available at:https://sol.sbc.org.br/
index.php/courb/article/view/2577.
Santin, P., Gubert, F. R., Fonseca, M., Munaretto, A., and
Silva, T. H. (2020). Characterization of public transit mo-
bility patterns of different economic classes. Sustainabil-
ity, 12(22). DOI: 10.3390/su12229603.
Santos, G., Gubert, F., Delgado, M., and Silva, T. (2024).
Redes de interesse: comparando o google places e
foursquare na captura da escolha de usuários por áreas
urbanas. In Anais do VIII Workshop de Computação Ur-
bana, pages 99–112, Porto Alegre, RS, Brasil. SBC. DOI:
10.5753/courb.2024.3248.
Senefonte, H. C. M., Delgado, M. R., Lüders, R., and Silva,
T. H. (2022). Predictour: Predicting mobility patterns of
tourists based on social media user’s profiles. IEEE Ac-
cess, 10:9257–9270. DOI: .
Shouji Du, Shihong Du, B. L. X. Z. and Zheng, Z.
(2020). Large-scale urban functional zone mapping
by integrating remote sensing images and open social
data. GIScience & Remote Sensing, 57(3):411–430. DOI:
10.1080/15481603.2020.1724707.
Silva, T. H., de Melo, P. O. S. V., Almeida, J. M., and
Loureiro, A. A. F. (2017a). Uma fotografia do instagram:
Caracterização e aplicação. Available at:http://143.54.
25.88/index.php/RB-RESD/article/view/74.
Silva, T. H., de Melo, P. O. V., Almeida, J. M., Musolesi,
Modeling Interest Networks in Urban Areas Santos et al., 2025
M., and Loureiro, A. A. (2017b). A large-scale study
of cultural differences using urban data about eating and
drinking preferences. Information Systems, 72(Supple-
ment C):95–116. DOI: 10.1016/j.is.2017.10.002.
Silva, T. H. and Fox, M. S. (2024). Integrating
social media data: Venues, groups and activities.
Expert Systems with Applications, 243:122902. DOI:
https://doi.org/10.1016/j.eswa.2023.122902.
Silva, T. H. and Silver, D. (2024). Using graph neu-
ral networks to predict local culture. Environment and
Planning B: Urban Analytics and City Science. DOI:
10.1177/23998083241262053.
Silva, T. H., Vaz de Melo, P. O. S., Almeida, J. M., Salles, J.,
and Loureiro, A. A. F. (2013). A comparison of foursquare
and instagram to the study of city dynamics and urban so-
cial behavior. In Proc. ACM SIGKDD Int. Workshop on
Urban Computing (UrbComp’13), Chicago, USA. DOI:
10.1145/2505821.2505836.
Silva, T. H., Viana, A. C., Benevenuto, F., Villas, L.,
Salles, J., Loureiro, A., and Quercia, D. (2019). Urban
computing leveraging location-based social network data:
A survey. ACM Computing Surveys, 52(1):1–39. DOI:
10.1145/3301284.
Silver, D. and Silva, T. H. (2023). Complex causal struc-
tures of neighbourhood change: Evidence from a func-
tionalist model and yelp data. Cities, 133:104130. DOI:
10.1016/j.cities.2022.104130.
Skora, L. E., Senefonte, H. C., Delgado, M. R., Lüders,
R., and Silva, T. H. (2022). Comparing global tourism
flows measured by official census and social sensing.
Online Social Networks and Media, 29:100204. DOI:
10.1016/j.osnem.2022.100204.
Traag, V. A., Waltman, L., and van Eck, N. J. (2019). From
Louvain to Leiden: guaranteeing well-connected commu-
nities. Scientific Reports, 9:5233. DOI: 10.1038/s41598-
019-41695-z.
Veiga, D. A. M., Frizzo, G. B., and Silva, T. H. (2019). Cross-
cultural study of tourists mobility using social media. In
Proceedings of the 25th Brazillian Symposium on Multi-
media and the Web, WebMedia ’19, page 313–316, New
York, NY, USA. Association for Computing Machinery.
DOI: 10.1145/3323503.3360620.
Wu, D. Q., Tan, J., Guo, F., Li, H., Chen, S., and Jiang, S.
(2020). Multi-Scale Identification of Urban Landscape
Structure Based on Two-Dimensional Wavelet Analysis:
The Case of Metropolitan Beijing, China. Ecological Com-
plexity, 43:100832. DOI: 10.1016/j.ecocom.2020.100832.
Ye, C., Zhang, F., Mu, L., Gao, Y., and Liu, Y. (2021). Ur-
ban function recognition by integrating social media and
street-level imagery. Environment and Planning B: Ur-
ban Analytics and City Science, 48(6):1430–1444. DOI:
10.1177/2399808320935467.
