ArticlePDF Available

The migrant perspective: Measuring migrants' movements and interests using geolocated tweets

November 2023
Population Space and Place 30(2)

DOI:10.1002/psp.2732

License
CC BY 4.0

Authors:

Johannes Mast

German Aerospace Center (DLR)

Marta Sapena

University of Valencia

Show all 5 authorsHide

Geolocated social media data hold a hitherto untapped potential for exploring the relationship between user mobility and their interests at a large scale. Using geolocated Twitter data from Nigeria, we provide a feasibility study that demonstrates how the linkage of (1) a trajectory analysis of Twitter users' geolocation and (2) natural language processing of Twitter users' text content can reveal information about the interests of migrants. After identifying migrants via a trajectory analysis, we train a language model to automatically detect the topics of the migrants' tweets. Biases of manual labelling are circumvented by learning community-defined topics from a Nigerian web forum. Results suggest that differences in users' mobility correlate with varying interests in several topics, most notably religion. We find that Twitter data can be a flexible source for exploring the link between users' mobility and interests in large-scale analyses of urban populations. The joint use of spatial techniques and text analysis enables migration researchers to (a) study migrant perspectives in greater detail than is possible with census data and (b) at a larger scale than is feasible with interviews. Thereby, it provides a valuable complement to interviews, surveys and censuses, and holds a large potential for further research.

shows the timeframe for the study, January 2015-December 2018. It was selected to allow for a consistent database with minimal temporal biases, defined by Olteanu et al. (2019) as systematic distortions across user populations or behaviours over time. Preliminary experiments suggested that the geotagging of places became widely used from mid-2014 onwards (Figure 1a). In 2019, Twitter made changes to the geotagging functionality (Kruspe et al., 2021) (Figure 1b). The use of Twitter by Nigerians spiked in 2020, fuelled by the #endsars protests (Ojedokun et al., 2021) (Figure 1c), while in 2021 the Nigerian government enacted a temporary ban on Twitter (Princewill et al., 2021) (Figure 1d). To prevent these events from influencing the study,

…

displays for each group of users the average share of users' tweets. Striking differences between user groups could be seen in the topics islam, religion and gaming.

…

Figures - uploaded by Johannes Mast

Content may be subject to copyright.

Content uploaded by Johannes Mast

Content may be subject to copyright.

Received: 21 November 2022

Accepted: 1 November 2023

DOI: 10.1002/psp.2732

RESEARCH ARTICLE

The migrant perspective: Measuring migrants' movements

and interests using geolocated tweets

Johannes Mast

|Marta Sapena

|Martin Mühlbauer

Carolin Biewer

|Hannes Taubenböck

1,3

German Aerospace Center (DLR), German

Remote Sensing Data Center (DFD),

Weßling, Germany

Department of English and American Studies,

Chair of English Linguistics, Julius‐

Maximilians‐Universität Würzburg,

Würzburg, Germany

Institute for Geography and Geology,

Julius‐Maximilians‐Universität Würzburg,

Würzburg, Germany

Correspondence

Johannes Mast, German Aerospace Center

(DLR), German Remote Sensing Data Center

(DFD), Germany, Münchener Str. 20,

Weßling 82234, Germany.

Email: johannes.mast@dlr.de

Funding information

European Union's Caroline Herschel

Framework Partnership Agreement on

Copernicus User Uptake (FPCUP),

Grant/Award Number: FPA275/G/GRO/

COPE/17/10042; Volkswagen Foundation,

Programme “Momentum—Funding for

Recently Tenured Professors”,

Grant/Award Number: 98662; German

Federal Ministry of Education and Research,

Grant/Award Number: 01LG2082C

Abstract

Geolocated social media data hold a hitherto untapped potential for exploring the

relationship between user mobility and their interests at a large scale. Using geolocated

Twitter data from Nigeria, we provide a feasibility study that demonstrates how the

linkage of (1) a trajectory analysis of Twitter users' geolocation and (2) natural language

processing of Twitter users' text content can reveal information about the interests of

migrants. After identifying migrants via a trajectory analysis, we train a language model to

automatically detect the topics of the migrants' tweets. Biases of manual labelling are

circumvented by learning community‐defined topics from a Nigerian web forum. Results

suggest that differences in users' mobility correlate with varying interests in several topics,

most notably religion. We find that Twitter data can be a flexible source for exploring the

link between users' mobility and interests in large‐scale analyses of urban populations.

The joint use of spatial techniques and text analysis enables migration researchers to (a)

study migrant perspectives in greater detail than is possible with census data and (b) at a

larger scale than is feasible with interviews. Thereby, it provides a valuable complement to

interviews, surveys and censuses, and holds a large potential for further research.

KEYWORDS

domain adaptation, human migration, mobility, NLP, social media, trajectories

1|INTRODUCTION

Human migration is a complex phenomenon, and its study is still

constrained by a lack of data (Kirchberger, 2021). Rather than being a

precisely defined process, migration is part of the spectrum of human

mobility patterns which range from occasional travel to permanent

relocation (Willekens et al., 2016). In a globalized world, migration is

intertwined with political, economic and cultural processes

(McAuliffe & Ruhs, 2018), but how these processes factor into

migrant decision‐making is still not sufficiently understood. Important

insights could be gained by better understanding the perspective of

individuals (McAuliffe et al., 2018), those who decide to move as well

as those who decide to stay (Schewel, 2020). Who are they, where

are they, and what matters to them? A data source which can keep up

with dynamic and international populations across borders, and

inform about their views and experiences in changing environments,

would be highly valuable (McAuliffe et al., 2018).

The same dynamism and variety of mobility patterns necessitate

that the data not only provide high coverage, but can also be flexibly

applied across space, time and topics. This is where traditional data

Popul Space Place. 2023;e2732. wileyonlinelibrary.com/journal/psp

1of17

https://doi.org/10.1002/psp.2732

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,

provided the original work is properly cited.

sources, such as ethnographies, surveys and censuses, reach their

limits (Rampazzo et al., 2021). Ethnographies and interviews provide

rich detail, but rely on small samples (Rowe et al., 2021). Longitudinal

surveys capture migration dynamics with detailed migration histories

but at high data collection costs and efforts (Fussell et al., 2014).

Demographic data from censuses or administrative sources can

provide high geographic detail for large parts of the population, but

are limited in detail, timeliness and coverage of migrants, and suffer

from heterogeneity in their definitions of migration (Rampazzo

et al., 2021; Spyratos et al., 2018; Willekens, 2019). Altogether, data

suffers from cost and rigidity, and there appears to be a gap between

detailed and large‐scale information sources.

These information gaps could be reduced by types of data which

have emerged from developments in information technologies

(McAuliffe et al., 2018; Reips & Buffardi, 2012): Microblogs like

Twitter provide geolocated social media (SM) data that is, to date,

freely and globally available. The combination of text data and

geolocation which these data offer enables the joint assessment of

the interests and the mobility of users and reveals pathways towards

a better understanding of migration.

2|BACKGROUND: ANALYSIS OF

MIGRANTS USING SM

In recent years, the number of people engaging with SM has grown

substantially, and SM platforms are now recognized as spaces for

socializing and reflecting on all aspects of everyday life (Townsend &

Wallace, 2016; Zhu et al., 2022). Consequently, “social sensing”data

have been proposed as a real‐time and inexpensive way to measure

social phenomena (Wang et al., 2019).

The opinions expressed on SM can inform about a wide variety

of topics (Wang et al., 2019), making the usage of user‐provided data

a scientific trend in many fields of research (Kounadi & Resch, 2018).

For migration studies, text‐based social networks are an interesting

data source, because they enable users to connect and exchange

knowledge over long distances (Dekker & Engbersen, 2014), and

have been considered central to international migration decision‐

making (Akanle et al., 2021). For qualitative studies, migration

researchers have long recognized SM's potential to provide a unique

insight into the interests and behaviours of migrants (Reips &

Buffardi, 2012). But could migrant interests also be studied at large

scales, using big data approaches?

The feasibility of large‐scale studies of SM users has been

demonstrated for a variety of text‐based SM platforms, such as

LinkedIn (Bastian et al., 2014), Facebook (Heidenreich et al., 2020)or

Twitter (Giachanou & Crestani, 2016). Widespread methods for the

analysis of such data include topic modelling (Calderón et al., 2020)

and sentiment analysis (Giachanou & Crestani, 2016). When it comes

to migration research, however, these studies have a blind spot.

Khatua and Nejdl (2021) found that, so far, SM studies mostly

explored public opinions about refugees and migrants (see, e.g.,

Heidenreich et al., 2020; Lee & Nerghes, 2018; Rowe et al., 2021),

while the first person‐perspectives of migrants have been neglected.

This is a missing link of no small importance (McAuliffe et al., 2018):

Subjective perceptions are key to migration decisions (Hoffmann

et al., 2021), even if they do not always match the objective reality. In

their aforementioned study, Khatua and Nejdl (2021) also demon-

strate the feasibility of analyzing views and struggles of migrants on

SM, but find that identifying migrants is challenging. Their approach

of relying on migrants explicitly referring to themselves as such can

detect only a miniscule, likely biased (Olteanu et al., 2019), subset of

migrants, and is thus not able to make visible the needs and desires of

the large number of migrants that are active on SM.

However, there are alternative methods which can be used to

identify migrants. Language can be used as a proxy (Lamanna

et al., 2018; Sîrbu et al., 2021), although limited by the dominance of

English as a lingua franca (Kim et al., 2020). And where SM contains

geoinformation, spatial analysis is a possibility. Various studies

(Armstrong et al., 2021; Blumenstock, 2012; Fiorio et al., 2017;

Gollin et al., 2021; Hawelka et al., 2014; Mazzoli et al., 2020;

Spyratos et al., 2018; Zagheni et al., 2014) show that the

identification of mobility patterns using geolocated SM data is

feasible. Armstrong et al. (2021) find that the users identified by their

method rarely represent migrants in the traditional sense but cover a

wide range of mobile users including business travellers, tourists or

global citizens. Therefore, SM data have the potential to be a

common ground for the analysis and comparison of a range of mobile

populations of various kinds which may not jointly appear in

traditional data sets.

However, thus far, the analysis of movements from the point of

view of geographical science and the analysis of texts from the point

of view of social science have remained apart. Recently, works by

Kim et al. (2020,2021,2022) have shown that migrants and natives

can be distinguished by relating their geolocation to that of their

friends (Kim et al., 2020), and characterized via the metadata and

hashtags associated with their tweets (Kim et al., 2021). This

demonstrates the potential of combining various facets of SM data,

and while Hashtags are powerful labels for trending topics, an even

greater wealth of information is contained in the main content of SM

posts. Therein, user‐created texts contain implicit and subjective

information about a wide variety of topics and the users' attitudes

towards them. Extracting this information from large data sets is only

feasible with automated methods though, which is challenging

because the features of informal Internet communication are

different from traditional written text and often contain typographi-

cal errors (Nguyen et al., 2020).

Over the past years, substantial advances have been made in the

field of natural language processing (NLP). Large general‐purpose

language models trained on very large corpora (Brown et al., 2020;

Devlin et al., 2018) now excel at a variety of tasks, including

multilabel text classification. This facilitates the automated extraction

of complex topics even from large corpora in informal language

(Kayastha et al., 2021). So far, the full potential of this has not been

tapped in the context of migration, perhaps due to the lack of suitable

training corpora or the aforementioned difficulty of identifying

2of17

MAST ET AL.

15448452, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/psp.2732 by Cochrane Germany, Wiley Online Library on [28/11/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

migrants. In this study, we demonstrate how these challenges can be

overcome.

Our overarching goal is to develop a methodology which can

harness geolocated SM data for migration research. Linking Nigerian

Twitter users' mobility to their interest in certain topics, we provide a

test case which explores the potentials and limitations of our method.

As an alternative to the recent approach by Kim et al. (2021), we aim

to provide a method that, given the availability of sufficiently rich

data, can be flexibly adapted to various mobility forms and platforms,

and captures general topics rather than specific hashtags, to match

the variety of mobility forms and online conversations.

Our approach is twofold:

First, we use the geolocation information to identify mobile users

via their trajectories. Second, we analyze the text content produced

by these users to identify their topic preferences and compare them

to those of stationary users.

The remainder of this paper is structured as follows. After an

outline of the setting and the data source of our study (Section 3.1),

we present the materials and methods used for the mobility analysis

(Section 3.2.1) and the text analysis (Section 3.2.2). The results are

likewise presented separately for the spatial analysis (Section 4.1) and

the text analysis (Section 4.2). The discussion (Section 5) reflects on

the key findings of the case study (Section 5.1) and discusses the

implications for policy (Section 5.2), closing with a reflection on the

limitations of the approach and ethical considerations (Section 5.3).

Conclusions are drawn in Section 6.

3|MATERIALS AND METHODS

3.1 |Study area and data source

Nigeria provides a relevant case study due to its cultural diversity,

pivotal position in West Africa, and rapidly increasing young

population (Central Intelligence Agency, 2022) with increasing

penetration of mobile phones (Forenbacher et al., 2019). Today, as

Kirwin and Anderson (2018) found, Nigerians comprise the vast

majority of people in West Africa who are motivated to migrate. SM

use is relatively low, but increasing. In January 2015, 13.6 million SM

users were estimated in Nigeria, a 7% share of the country's total

population (Kemp, 2015). In January 2019, that number had risen to

24 million, or 12% of the population (Kemp, 2019). The users are

predominantly young, with 68% being between 18 and 34 years old

(Kemp, 2019). Consequently, SM users are not representative of the

population as a whole. However, they still constitute a userbase of

large size whose data can possibly complement existing data or even

give new insights into migration.

In this study, Twitter was used as a data source for the following

reasons:

(1) Joint geolocation and text data: Tweets are text data that are also

sometimes tagged with locations (place geolocation) or geocoor-

dinates (precise geolocation). Geotags link the physical space

with the virtual space of online conversations.

(2) Rich metadata: Besides geolocation, tweets are accompanied by a

variety of metadata which aid in the interpretation of the data

and the detection of automated accounts.

(3) Accessibility: Twitter's accessibility has made it the most studied

SM platform (Williams et al., 2017). An Application Programming

Interface (API) provides access to historical data, although with

restrictions to query volume and rates (Twitter Inc., 2022a). For

these reasons, Twitter is a preferred data source for spatial

analyses, despite estimations that only 1.1% of the adult

population can be reached with adverts on Twitter, compared

with 18.0% for Facebook (Kemp, 2019).

Figure 1shows the timeframe for the study, January

2015–December 2018. It was selected to allow for a consistent

database with minimal temporal biases, defined by Olteanu et al.

(2019) as systematic distortions across user populations or beha-

viours over time. Preliminary experiments suggested that the

geotagging of places became widely used from mid‐2014 onwards

(Figure 1a). In 2019, Twitter made changes to the geotagging

functionality (Kruspe et al., 2021) (Figure 1b). The use of Twitter by

Nigerians spiked in 2020, fuelled by the #endsars protests (Ojedokun

et al., 2021) (Figure 1c), while in 2021 the Nigerian government

enacted a temporary ban on Twitter (Princewill et al., 2021)

(Figure 1d). To prevent these events from influencing the study,

FIGURE 1 Sampled count of geolocated tweets from Nigeria between 2011 and 2022, by geolocation type. (a) First appearance of

substantial numbers of geotagged Tweets in Nigeria; (b) change to Twitter geolocation functionality (Kruspe et al., 2021); (c) #endsars protests in

which Twitter played a significant role (Ojedokun et al., 2021); (d) temporary ban of Twitter by the Nigerian government (Princewill et al., 2021).

MAST ET AL.

3of17

the 4‐year period 2015–2018, for which a stable level of general

activity and geotagging activity can be found, was selected as the

study's timeframe.

As an ancillary data source, data from Nairaland.com, a Nigerian

web forum was used. For a description of Nairaland (NL) we refer to

Supporting Information Appendix B.

3.2 |Methods

The methodology follows a complementary workflow of spatial

analysis and text analysis. On the basis of a source set of Nigerian

tweets (S

NGA

), which is expanded in geographical scope to S

GEO

, the

spatial analysis identifies several groups of users with distinct

mobility characteristics (U

and U

in Figure 2). For these

mobile people, a new set of tweets (S

MPL

) is acquired to analyze their

topic interests. Figure 2illustrates how outputs of the spatial analysis

are being used to guide the data acquisition of the text analysis.

3.2.1 |Spatial analysis

For the timeframe, all geolocated tweets from Nigeria were queried

using the Twitter API. From the over 28.5 M tweets collected (data

set S

NGA

, Table 1), unique users were identified with the aim to

analyze their movement. To exclude automated accounts, a series of

filters was applied following previous studies on trajectories by Hübl

et al. (2017) and Petutschnig et al. (2020): First, accounts that

frequently (for more than 25% of tweets) exceeded the speed of

150 km/h between two sequential tweet locations were excluded.

Besides filtering automated accounts, this speed‐based filter ex-

cluded users who use the geotagging feature consistently in other

ways than to refer to their own location (e.g., referring to a place they

intend to visit, or the location of an event they comment on). Second,

accounts that post more than 150 geolocated tweets in a single day

(Hübl et al., 2017), or more than 15 on average per day (Petutschnig

et al., 2020), were excluded. As a final measure to improve

consistency, only users who were created before the start of the

timeframe were selected. For these remaining 116,670 users (U

NGA

all tweets from outside Nigeria for the timeframe were acquired to

complete their timelines of geolocated tweets (data set S

GEO

Table 1).

For the mobility analysis, as small‐scale mobility was not of

interest in the study, the tweets from S

GEO

were spatially aggregated

to spatial units (SUs) based on the centroid of the place tagged in the

tweet. This further serves to anonymize the data. To ensure that the

size of the SU did not determine the results, a sensitivity analysis was

performed and confirmed that the results were robust to the size of

SU influence on the result excepting extreme values. SU of a 30‐km

radius was used for the remainder of the study. For more information

on how the SU was derived, see Supporting Information Appendix A.

The geolocated timelines S

GEO

contain location histories for

116,670 users. To identify migration events within these histories,

FIGURE 2 Workflow of the study. (Top) Spatial analysis. (Bottom) Text analysis. S

NGA

GEO

and S

MPL

are Twitter data sets. NL is a Nairaland

data set (see Section 3.2.2). M

OVE

and M

INT

are movements. U

NGA

is a set of users, of which U

and U

none

are subsets (groups) of users.

GYR

is the radius of gyration.

TABLE 1 Twitter data sets used in the study.

Data set

Number of

tweets

Number

of users Geolocation Purpose

NGA

27,818,148 325,061 Complete Identification of

users U

NGA

GEO

25,021,363 114,278 Complete Mobility

analysis of

users U

NGA

MPL

2,629,812 9301 Partly Text analysis of

users U

and U

Note:S

GEO

, global data set for selected users U

NGA

MPL

, sampled data set

for analysis of mobile populations; S

NGA

, initial source data set of Nigeria.

4of17

MAST ET AL.

the approach developed by Chi et al. (2020) was applied. It can be

flexibly tuned to detect migration in various forms and has been

calibrated and validated on several data sets, Twitter among them. In

this three‐step approach: (1) contiguous segments are identified, (2)

consecutive segments in the same location are merged and (3)

overlaps are removed. Compared with Chi et al. (2020), one of

several overlapping segments was allowed to persist if they

contained more than 70% of all tweets in the overlapping period.

The results are continuous, nonoverlapping segments that indicate a

user's presence at a certain location over a certain time (Figure 3).

Segments qualified as residences if their length exceeded a

minimum time (ResLength). Minimum residency length is part of many

definitions of migration but no single definition is universally

accepted (Kirchberger, 2021). To demonstrate the flexibility of the

approach with regard to migration concepts, two types of movement

and four groups of users were distinguished (see Figure 2): The first

type of movement was defined as international migration (M

INT

which occurred between two residences of at least 90 days

ResLength, only one of which was within Nigeria. The 90‐day window

corresponds to the commonly used distinction between visitors and

longer terms stays as codified in visa‐waiver programmes (Armstrong

et al., 2021). It corresponds to the UN definitions of “short‐term

migrant”and “long‐term migrant”which refer to residence periods of

3–12 months, and longer than 12 months, respectively (United

Nations Department of Economic and Social Affairs, 1998). The

second type of movement was defined as overall mobility (M

OVE

which occurred between any two different residences of 30 days

ResLength. It was intended to capture a wider range of movements

that could also include short‐term mobility.

On the basis of these two types of movements, four groups of

users were distinguished: international, mobile, stationary and others.

First, international migrants (U

) were users who displayed at least

FIGURE 3 Illustration of the migration detection method on six fictitious user timelines. Consecutive tweets of sufficient density get joined

to segments. The succession of segments at different locations is counted as a movement or migration. According to the migrations, users are

categorized as either U

or U

none

. NG, Nigeria; US, the United States.

MAST ET AL.

5of17

one M

INT

. Second, mobile users (U

) displayed at least one M

OVE

within Nigeria and no M

INT

. Third, stationary users (U

) were those

which had several detected residence periods of at least 30 days all at

the same location in Nigeria. To prevent inactivity being mistaken for

spatial continuity (Schewel, 2020), stationary users had to possess at

least two residence periods in the same Nigerian SU (cf. Figure 3e,f),

and further fulfil the requirement of having a higher number of total

tweets than the lowest number of tweets exhibited by any U

or U

(19 tweets). Lastly, other users U

none

fulfilled none of the other

groups' criteria and were not of interest to this study.

As a final analysis step, the movements from one residence to

another were aggregated to mobility flows between the SU.

Assigning users and movements to discrete categories allows

testing for a priori definitions of migration, but it is not the only way

to make mobility tangible. Location histories alternatively allow for

the quantification of user mobility on a continuous scale. One

possible measure is the radius of gyration (R

GYR

), the mean distance

of a user's geolocated tweets from their collective centroid (see

Zagheni et al., 2014 for an example). To test this alternative method

of capturing mobility, R

GYR

was calculated for U

and U

3.2.2 |Text analysis

The objective of the study was to identify whether the previously

identified differences in mobility also extend to differences in topic

interests of the particular user groups, expressed by the tweeting

about certain topics.

The geolocated tweet timelines S

GEO

, used in Section 3.2.1 to

detect mobility, were not suitable for the analysis of topics, as they

only contained geolocated tweets. These are likely biased towards

certain topics that have a strong spatial component (e.g., travel)or

ties to particular locations (e.g., sports). To overcome this limitation, a

new set of tweets S

MPL

was queried without a requirement for

geolocation.

A technique of random sampling, stratified by users, was applied to

download a representative selection of tweets from all users who were in

the previous step classified as either international, mobile or stationary.

This sampling was necessary for a couple of reasons: (1) Acquiring the

users' complete tweet histories for the time between 2015 and 2019 is

theoretically possible, but practically unfeasible due to the high number of

tweets. And (2), a complete data set would be highly biased towards more

active users (Li et al., 2013; Zagheni et al., 2014).

Random sampling of tweets from users' tweet histories is not

offered by the Twitter API. As a heuristic alternative, a pseudo‐

random stratified sampling approach was implemented by sampling

48 randomly spaced 7‐day intervals throughout the timeframe, and

acquiring for every user up to 10 tweets from within each of these

intervals. Tweets from highly topic‐specific sources or automated

platforms were discarded.

The resulting data set S

MPL

comprises 2,731,483 tweets from

9672 users. Basic summary statistics across user groups are provided

in Table 2.

For the classification of topics within the tweets, we apply a

machine‐learning model. To train this model, labelled training data are

necessary. As an alternative to manual labelling of tweets, implicitly

labelled training data from the Nigerian web forum NL were acquired.

NL consists of many subforums dedicated to certain topics,

like, politics,sports and travel. By choosing the appropriate subforum

for their comments, users are implicitly labelling their own texts.

Therefore, such hierarchically structured web forums can be seen as

a labelling environment in which labels are offered top‐down and

assigned by the web community bottom‐up, with certain community

members having more control over the labelling than the majority.

Compared with a manual labelling of posts or definition of topic‐

specific keywords by experts, automatically learning the labels from

the online community is more likely to reflect the themes discussed in

online spaces and the language that is being used to discuss them.

While this approach does not avoid subjectivity or normativity, it

makes for a better fit to the microblog format. In total, 2,091,491 NL

comments spread over 40 topics were acquired using a web‐scraping

approach. Of the NL comments, 80% were used for training, 10% for

validation and 10% for testing the model and calculating the topic‐

specific accuracies for the plausibility check. For a more detailed

description of the NL data set, see Supporting Information

Appendix B.

To analyze the interests of Twitter users, a state‐of‐the‐art

language model was trained on the NL data set and then used to

assign topic labels to all the tweets in S

MPL

. The feasibility of a similar

approach has been demonstrated by Fiallos and Jimenes (2019) who

labelled the interests of Twitter users' based on a classifier trained on

Reddit comments.

BERTweet (Nguyen et al., 2020) was used as the backbone of the

model. It is a language model based on the transformer architecture

(Vaswani et al., 2017) that is pretrained on a large English Twitter

corpus. This backbone converts the text into vectorized embeddings

TABLE 2 Statistics of users over S

MPL

Number of users

Number of tweets R

GYR

(km)

Minimum Maximum Mean SD Minimum Maximum Mean SD

Stationary U

3748 13 480 282.4 128.7 0.6 5879.7 413.2 714.7

Mobile U

4655 1 480 276.4 128.0 1.5 8204.4 864.4 1258.9

International U

898 22 480 316.9 117.9 66.4 10,045.5 2478.4 1463.7

Note:R

GYR

, radius of gyration; SD, standard deviation; U

, stationary users; U

, mobile users; U

, international migrants.

6of17

MAST ET AL.

which feed into a fully connected classification layer. The entire

network, including the transformer backbone, was then fine‐tuned to

the task of predicting the topic for each NL comment. Afterwards,

the trained classification model was applied to the Twitter data set

MPL

to predict the topics for each tweet.

This approach is one of unsupervised domain adaptation (see

Ramponi & Plank, 2020, for an overview) in which S

MPL

constitutes

the target domain (D

) and NL constitutes the source domain (D

Despite many similarities between NL and the Nigerian Twitter-

sphere, there remains a domain shift between D

and D

due to

differences in userbase, moderation or user options. To reduce this

shift, tweets and NL comments were preprocessed identically,

removing obviously platform‐specific features (such as the retweet

marker “RT”) and masking unique items, such as URLs or phone

numbers. As in the original training of BERTweet, comments shorter

than 10 tokens were excluded. On grounds of representativeness,

the small proportion of non‐English tweets and NL comments were

not excluded, although their classification likely added challenge of

the task. To the model, two further adjustments were made

compared with the reference implementation. First, to account for

unequal representation of topics in our NL data set, cross‐entropy

loss with class weights was used (Torch Contributors, 2019). Second,

a domain‐adversarial domain adaptation approach was applied during

training (see Ganin et al., 2016) to promote the learning of domain‐

invariant features. It was also found to serve as a regularization

technique that reduces overfitting and improves accuracy.

The model was trained for five epochs and reached a validation

accuracy of 42.5% on withheld NL validation data, an acceptable

result considering multiclass prediction on 40 classes is a much more

difficult task than single‐label prediction (Quercia et al., 2012). The

accuracy is comparable to the 42.1% accuracy of the best‐performing

model in the TREC‐IS 2018 challenge involving 25 classes of disaster‐

related tweets (McCreadie et al., 2019).

Within each predicted topic, the significance of differences

between the groups U

and U

was determined via the

Kruskal–Wallis rank sum test.

Due to the lack of labelled tweets, the model's performance on

tweets cannot be gauged with traditional validation metrics. As an

alternative, a threefold credibility check was applied to each topic.

First, the topic‐specific F1‐score was calculated on the predic-

tions on a withheld NL test set that comprised 10% of all NL

comments. If a topic could be accurately classified in D

, the same

would be true in the D

. The F1‐score was calculated as follows:

Precision TP

TP FP

=+,

Recall TP

TP FN

=+,

Fscore Precision Recall

Precison Recall

2× ×

TP FP

where is the true positives, the false positives

and the false negatives.

Second, characteristic weekday activity patterns were checked

for similarity. When the topic exhibited a similar pattern in the tweets

as in the NL data set, as quantified by the Pearson correlation

coefficient (COR) between their activity shares over weekdays

(Leydesdorff, 2005), the model was able to recognize a particular

topic in the Twitter domain. This check was not applied for topics

where a coefficient of variation (CV) lower than 0.05 indicated the

absence of any noticeable weekly pattern (e.g., gaming in Figure 4).

Third, ancillary data sets of NL comments NL

NEW

and of tweets

NEW

were acquired from a more recent period (2021–2022). From

NEW

keywords were extracted that did not appear in the older NL

data set. These keywords were assumed to represent novel concepts

(e.g., “covid19”). If they appeared in similar topics in S

NEW

as in

NEW

, it supported the assumption that the Twitter topics matched

NL topics. To quantify this similarity for each topic, the cosine

similarity (Egghe & Leydesdorff, 2009) between NL

NEW

and S

NEW

was calculated using the topic's shares of the usage of the

aforementioned keywords as features.

FIGURE 4 Weekly patterns and correlations of example topics. (Green) True labels of all NL posts. (Orange) Predictions on NL test data

set. (Blue) Predictions on Twitter data set. y‐Axis: Proportion of the weekly maximum by topic and data set. COR, Pearson correlation between

NL and Twitter patterns; CV, coefficient of variation; NL, Nairaland.

MAST ET AL.

7of17

To condense this three‐pronged check into a tangible measure, the

predicted topics were grouped into four discrete grades of credibility

from A (most credible) to D (least credible), based on whether they failed

to meet a grade‐specific threshold in all three checks. Thresholds of 0.5,

0.4, 0.3 and −1 were used for grades A, B, C and D, respectively.

To complement the quantitative metrics of plausibility, we

perform a qualitative reading of sampled tweets on those topics for

which we find the strongest differences between user groups, as well

as the topic which likely holds the most migration‐related informa-

tion. Our goal is to better understand how much variation exists in

the content that is algorithmically bundled into a topic. For each

topic, 100 tweets were sampled for each user type U

and U

for a total of 900 tweets across three topics. Each tweet was

assigned a code, based only on its raw text and the content of any

linked web pages or images. Following grounded theory methodology

(Charmaz, 2006), codes were not defined by preconceived hypothe-

ses but developed from the observations in the data in an exploratory

manner, although with a focus on mobility‐related information.

Consequently, the goal was not an overarching framework of themes,

emotions or styles across all topics but a specific set of codes for each

topic which best represents the variation within it.

4|RESULTS

4.1 |Results of the spatial analysis

Figure 5displays the migration flows between the most connected

places for overall movements (M

OVE

) and international migrations

INT

). Of 10,101 M

OVE

, 50.9% were within Nigeria, 18.8% were

outgoing, 18.4% were incoming and 11.4% were between countries

outside Nigeria.

Of the 1057 M

INT

, 48.0% were incoming to Nigeria while 52.0%

were outgoing. The capital Abuja and the largest city Lagos are

measured with the strongest international ties, most connecting to

the United States and the United Kingdom.

As for the users, the number of mobile users (U

) was 4794, of

which 3229 (67.4%) moved purely within Nigeria. For roughly a third

(35.3%) of the U

, we observed multiple M

OVE

in the timespan.

The international class U

consisted of 926 users. For 111

(12.0%) of these, we observed multiple international migrations in the

timespan.

This means that we measured some mobility for 5720 (4.9%) of

our 116,670 analyzed users.

Of 3857 stationary users (U

), most resided in Lagos (49.2%), Abuja

(20.0%), Kaduna (3.5%), Ibadan (3.3%) and Port Harcourt (3.2%).

The gyration radius is right skewed with a median of 223 km and

a mean of 840 km. Interestingly, the distribution of its decadic

logarithm is clearly bimodal, with two peaks corresponding to radii of

approximately 200 and 2200 km (Figure 6).

4.2 |Results of the text analysis

The credibility of the domain adaptation varied strongly by topic. In

this section, we provide results for credibility grades A, B and C. The

complete results of the credibility check and the topic analysis for all

topics are provided in Appendices Cand D.

FIGURE 5 (Left) M

OVE

s between SU with at least 250 M

OVE

s. (Right) M

INT

s between SU with at least 25 M

INT

s. SU, spatial unit.

8of17

MAST ET AL.

Figure 7displays for each group of users the average share of

users' tweets. Striking differences between user groups could be

seen in the topics islam,religion and gaming.

Islam has been assigned to 3.86% of U

tweets, but only 3.68%

of U

and 3.31% of U

tweets. The reverse pattern was found for

religion, with 2.64%, 3.44% and 3.77% for U

and U

respectively. Gaming made up 3.05% of U

tweets, but only 2.54%

of U

and 2.11% of U

tweets. The topic properties seemed to be of

interest for U

while music‐radio was of interest to U

. Notably, U

intended to represent mobile users, had lower shares in travel than

either U

and U

For the topics gaming and religion, the strongest differences

in interest were found across user groups. They, along with topic

travel, were qualitatively evaluated. The codes, displayed in

Figure 8and described in Supporting Information Appendix C,give

indication about the content of the topics. Gaming, more popular

with international migrants, was found to include a mix of

entertainment‐related subtopics, most commonly football, but also

other sports, games and music. Public conversation about aspects of

public life (celebrities, fashion), private life (relationships and

pets) and politics were also common, making up around 45% of

the sampled tweets. Religion, on the other hand, was thematically

homogeneous, with around 18% of the tweets not clearly related to

religion. It was found that the religious discourse took a variety of

forms, ranging from bible quotes to proclamations of devotion or

prayers to discussions about matters of politics or proper behaviour.

Dominant in the topic travel was various forms of mobility‐related

information at various spatial scales (from local traffic updates by

users to international news by journalists) as well as news,

information and discussions about places or events with an explicit

spatial location. Comments on the state of transport infrastructure

and transport service providers were predominantly critical. Inten-

tion to travel or migrate was only occasionally (2%) expressed, while

sharing of personal travel experiences was more common (~13%).

Around 25% of tweets were not clearly mobility related.

As described in Section 3.2, the gyration radius R

GYR

provides a

continuous measure of mobility, which allows the exploration of the

relationship between mobility and interest without grouping users.

Due to the large number (n= 9672) of users, locally estimated

scatterplot smoothing (LOESS, Cleveland & Loader, 1996) was

applied to aid exploration of patterns in the relationship from a large

number of tweets (Figure 9). These patterns suggested that increased

mobility might correlate with lower interest in religion, as was

previously observed in the groupwise topic interest (Figure 7). The

inverse was true for gaming and particularly travel, where, past a

certain point, the interest in the topic seemed to be logarithmically

related to R

GYR

. For islam and properties, the curves suggested a more

complex and multimodal relationship.

FIGURE 6 Histogram of the decadic logarithm of gyration radii.

FIGURE 7 Averaged topic interest of various user groups. Topics are graded by credibility into grades (a), (b) and (c). Significance of

differences between groups was determined via Kruskal–Wallis rank sum test.

MAST ET AL.

9of17

5|DISCUSSION

5.1 |Key findings of the case study

The detected mobility flows suggest that most international migrations of

Nigerian Twitter users connect to large urban centres and support the

findings by Kirwin and Anderson (2018)thatamongNigerianurban

residents, Lagosians are more likely to want to migrate abroad. Besides

Lagos, the smaller Abuja has almost comparable international ties, likely

due to its unique status as the state's capital. We found the strongest

international connections to the United States and the United Kingdom,

which agrees with polls conducted at the start of our study period (NOI

Polls Ltd., 2015). However, the well documented movements of refugees

and internally displaced persons (IDPs) in the north of Nigeria

(International Organization for Migration [IOM] & Displacement Tracking

Matrix [DTM], 2022) were not reflected in our findings, confirming that

FIGURE 8 Frequency of codes in the qualitative reading of topics gaming, religion and travel. For each topic, 100 tweets were randomly

sampled per user group (U

and U

). Numbers on the bars refer to the tweet count.

FIGURE 9 Locally estimated scatterplot smoothing curve of topic interest versus the decadic logarithm of user gyration radius. Due to the

large number of points (9672) only the fitted curves are shown. Note that the x‐axis is logarithmic. The shaded areas between the dashed lines

indicate the span of the 95% confidence interval.

10 of 17

MAST ET AL.

refugees and IDPs are likely less or not active on Twitter, as discussed by

Petutschnig et al. (2020).

Our results suggest that international migrants have, on average,

a higher interest in the topic islam than stationary users, who in turn

seem to have more interest on the topic religion. This is in contrast

with previous work by Bloch et al. (2015) and a survey conducted by

Kirwin and Anderson (2018), which found that within Nigeria,

Muslims are less likely to want to migrate abroad. It is possible that

the relationship between religion, islam and mobility is more complex

than was previously assumed, at least for the demographic of Twitter

users (see Section 5.3). This view is also supported by the LOESS

regression for the topic islam, which painted a more complex picture

than the categorical results suggested. International migrants

displayed significantly higher interest in the topic gaming, which

consisted of a surprising variety of entertainment‐related subtopics,

such as music, celebrities and above all football. These may not, by

themselves, be significant migration factors. But football is of great

importance to Nigerian culture, and the attachment to European

football brands creates strong international ties with consequences

both positive and troubling (Igwe et al., 2021). The topic travel,

selected because its label suggests relevance to mobility studies,

curiously was observed with the highest popularity with stationary

users, albeit by only a small margin. As we found in our qualitative

reading, the topic comprises a large variety of mobility‐related

themes, conversations and information, from a local to global scope.

Stationary and mobile users alike frequently commented on

inadequate transport infrastructure, traffic obstructions or conges-

tion, which are serious problems in Nigerian cities (Afolabi et al., 2017)

and have a wide range of negative impacts on people and the

economy (Economic Intelligence Unit, M. of E. P. & B., 2013). For

international migrants, on the other hand, we found that reports of

their own journeys and specific locations prevail. While these insights

are based on a relatively small sample of 900 tweets, they hint at the

type of insight that can be gained about the differences and

similarities between international migrants and stationary people

using such an approach as presented in this paper. In conjunction

with our quantitative results, we found indications that mobility is of

importance even for people who would be, at a large scale,

considered stationary. At a methodological level, we found that our

approach to identifying general topics accomplishes its inclusivity at

the cost of masking the complexity of information contained within.

This complexity can appear as diversity in subtopics, as in the

example of gaming. Or, as in religion, as a variety of conversation

types (e.g., advice, discourse, debate, prayer, quote). Or as in travel,

the information can vary in spatial and temporal scope, from ad‐hoc

traffic updates by local users to professional news reporting about

developments in international politics. Therefore, statistical analysis

of any topic should always be accompanied by in‐depth reading of at

least a sample of tweets. Future studies could aim at digging deeper

via a more detailed inspection of tweet messages in a conversational

context, considering the line of argumentation in the tweets, and

including statistics about linguistic features and metadata in a

multimodal approach.

5.2 |Policy implications

We find that our results are plausible and conclude that with SM data

and our methodological approach, it becomes possible to capture

issues that preoccupy mobile users at a general level. Thus, we offer

means to approximate the migrants' perspective, and the data allow a

detailed evaluation of the textual content based on this. Compared

with the approach by Kim et al. (2021), the approach presented in this

paper is more demanding (in terms of data) but also versatile in the

classification of mobility and content. For researchers, this grants

flexibility in the criteria they use to identify migrants, enables the

combination of multiple criteria in an ensemble (as proposed by

Johnson et al. (2016), and could contribute to a better understanding

of the link between short‐term mobility and long‐term migration

(Mau et al., 2015). Altogether our presented approach is likely less

suited to generating and forecasting demographic data and more

useful to studies of public opinion and research in the social science

domain, where it can serve as a timely intermediate between official

statistics and in‐depth qualitative research. We believe it will be

particularly useful for comparative studies of international migration,

where the granularity and extensive coverage of the geolocated SM

data provide a clear advantage over traditional data sources (Bosco

et al., 2022).

Our approach provides an overview of migrants' general

interests but naturally is not very detailed on any specific issue.

More detailed and qualitative approaches, which focus on specific

topics that are clearly defined by the researcher via keywords and

hashtags, are going to remain vital. Using different approaches in

conjunction, we believe, is a way by which researchers can gain a

more complete and less biased picture of migrants.

However, understanding what occupies migrants is not merely of

academic interest, but beneficial to the management of migration and

its impacts. As Kim et al. (2022) show, topics can be the key to

understanding attachment to places of origin or destination. And this

can result in substantial economic impact: In Nigeria, remittances,

given by ethnic Nigerians in diaspora, constitute a significant source

of foreign currency with a $20.9 billion inflow in 2022 (World

Bank, 2022). Neither this inflow nor its benefits should be taken for

granted, however, as Didia and Tahir (2022) found remittances do not

automatically enhance economic growth, with a large portion of

remittances going towards consumption and social insurance rather

than investment. They recommend that by better understanding the

diaspora and earning their trust, remittance behaviour could be

improved and channelled towards economic growth. Knowledge

from SM could support this by informing initiatives like the Nigerians

in Diaspora Commission (NiDCOM, 2021) about ways to connect

with the diaspora and learn about topics in which they might be

interested to invest in. Our findings indicate that compared with

stationary Nigerians, internationally mobile Nigerians engage more in

the public discourse on sports and other entertainment while

engaging less in discussions about—or public displays of—faith.

Engaging the diaspora on topics that are suited for online discussion

can be an important step to connect with the users on issues that

MAST ET AL.

11 of 17

matter to them, and possibly encourage them to invest remittances

on projects related to these issues. On the other end of the mobility

spectrum, those SM users who would prefer to remain in their

communities can also benefit from having their collective voices

heard and understood. Policy makers can engage and support them in

issues that matter to them, in the long run, also increasing the

resilience of the communities of stationary Nigerians.

Beyond the tangible benefits, we hope that the study of

migrants' interactions on everyday issues can preserve the human

perspective in the migration discourse and help migrants be seen as

more than mere atomic parts of a highly political issue. In our case

study we discovered a personal side of migrants, whose interactions

are also on nonpolitical topics, such as relationships, sports or films.

We hope that this can help build empathy and understanding, which

will indirectly benefit all aspects of migration management.

5.3 |Limitations and ethical considerations

Despite the proven capability of thepresentedapproach,weacknowl-

edge several limitations. We demonstrated that different forms of

mobility can be identified within the same data set by varying the

parameters of the algorithm. However, long information gaps between

identified residence segments, resulting from irregular tweet activity of

users, make it practically impossible to precisely define the residence

length and the number of relocations for most users. The identification of

seasonal migration patterns is greatly hampered by this data irregularity.

Simpler migration patterns are less reliant on the continuity of tweet

timelines and can be confirmed or rejected for more users. Consequently,

while the distinction of mobility at different rates (e.g., weekly, monthly,

seasonal)isonlyfeasibleforfewveryactiveusers,thedifferentiation

between mobile and stationary users can be made with a fair degree of

confidence for most users.

Similar considerations apply in the spatial dimension. A practical

constraint to spatial granularity is posed by the relatively small

number of tweets with precision at a point or neighbourhood level

(Kruspe et al., 2021). At the city scale, however, we can confirm the

robustness of our approach towards the size of the SU with a

sensitivity check (see Supporting Information Appendix A). We

conclude that the data are well suited for migration studies at

intercity, national and international scales and over long timespans.

The LOESS regression of mobility versus topic interest produces

intriguing patterns that paint, for some topics, a different picture than

the results by groups. While the method is exploratory, it illustrates

that, unlike traditional census or survey data, geolocated tweets offer

the opportunity to capture mobility without a priori definitions of

migration. Another promising use of continuous mobility information

is the identification of returners, which Pappalardo et al. (2015)

identified as people for whom recurrent movement constitutes a

large part of their mobility and are clearly distinct from explorers who

spread their movement over a larger number of locations.

Altogether, we find that geolocated SM data support studies

using a wide range of mobility concepts and can also theoretically be

implemented in a large variety of settings. However, there are

practical limitations: Changes in Twitter's policy, changes in user

activity and governmental restrictions affect the stability of the data

basis over time. Across space, the availability of geolocated tweets is

likewise not homogeneous. The population's affinity for geotagging, a

requirement for the analysis of mobility, varies across user types and

across countries (Huang & Carley, 2019), as does mobile phone

penetration (Gollin et al., 2021). Thus, while our approach theoreti-

cally supports many settings, we highly recommend that preliminary

experiments confirm the existence of a suitable data basis for each

particular application. More than a matter of technical viability, this is

a matter of representativeness.

It is generally agreed that Twitter users are not a representative

sample of the population (Spyratos et al., 2018; Taubenböck

et al., 2018). Despite the difficulty in assessing specific biases (Wang

et al., 2019), it is commonly asserted that SM platforms cater

especially to a young and urban population (Gollin et al., 2021;

Hughes et al., 2016). Malik et al. (2015) found for the US that a bias in

geotag usage further limits the representativeness of findings. This

limitation naturally extends to trajectory‐based analyses: According

to Armstrong et al. (2021) the forms of mobility detected by

Canadian tweets correspond less to migrants than to business

travellers or transnationals. Our own results support the finding by

Petutschnig et al. (2020) that Twitter is ill‐suited to inform about rural

migrations and refugee movements. On the flipside, it is suitable for

the study of transnationals, business travellers and urban–urban

migrants—forms of mobility that is widespread, but underrepresented

in research (Armstrong et al., 2021). On the basis of continued

urbanization and improved access to telecommunication in many

countries across the globe, we expect an increase in the number of

people which can be mapped with SM approaches.

We now turn from general demographics to a group of accounts

that are not controlled by a real person, but by some form of

automated algorithm (i.e., bot). The challenge of detecting and

filtering bots on Twitter has received considerable attention

(Efthimion et al., 2018; Orabi et al., 2020; Subrahmanian et al., 2016).

While not all types of automation are malicious (Orabi et al., 2020),

confusing bots with actual users will distort and discredit any analysis

intended to study human behaviour. In addition to simple initial filters

(see Section 3.2), we applied a posterior bot check. We found that

both (A) the migration detection algorithm (see Section 3.2) and (B)

the selection of general sources (see Section 3.2.2) were effective,

albeit imperfect barriers to nonmalicious automation, and conclude

that the influence of bots on our study is small.

We addressed the lack of labelled validation data by applying a

threefold credibility check which supports the interpretation of the

results by quantifying credibility. Of 40 considered topics, 30 do not

reach our standards across all checks. Rather than outright dismissal,

closer investigation of these topics might lead to further insights. The

topic sports illustrates this clearly: While the temporal patterns

between NL and Twitter show only a weak correlation, we observe a

high cosine similarity and classification accuracy. It is certainly

possible that this indicates an unsuccessful adaptation of sports from

12 of 17

MAST ET AL.

NL to Twitter. But an alternative, much more interesting explanation

is that the mismatch in temporal patterns results from the different

ways in which the two media platforms are used to discuss the

topic—perhaps Twitter lends itself to exchange about momentary

events, like, during a game, while web forums are used for predictions

about games or general debates. In‐depth analysis of specific (sports)

events could shed light on this question.

An important aspect of our approach is that the labels are not

predefined by the researcher but rather, in an implicit manner, by the

users. We find that this can lead to perceived misalignment between

content and label. For instance, after qualitatively analyzing the topic of

gaming and finding that it includes sports, music and film in addition to

video games and board games, we believe that the label “entertainment”

might be more accurate. Changing the name of the label is always an

option, of course, but risks imposing the researcher's biases about

semantics and related concepts. We refrain from any relabelling in this

study because we allow for the possibility that Nigerians use English in

different ways. Nevertheless, we believeitisavalidoptionwhenusedin

conjunction with a qualitative reading.

Comparison of the interests between topics is discouraged due to

substantial differences between precision and recall. Where the precision

is lower than the recall, the number of posts has been overestimated, and

vice versa. But, assuming roughly similar error rates between groups of

users—something we cannot validate in our current setup—even the

relatively low validation accuracy of 42.5% merely creates noise that

diminishes differences between user groups, but does not distort them

(cf. Figure 3, where the weekly patterns of jobs and religion are

weakened, but not completely eliminated in the Twitter data). Not only

does this mean that comparisons between groups are still credible but

also that differences between groups are likely even stronger in reality

than they appear in our results. Many errors are due to confusion

between similar and overlapping topics, such as phones and phone‐ads.

WhileweassumethattheseNLsubforumsareseparateforgoodreason,

if a qualitative reading reveals that distinguishing such topics does not

provide any relevant insight for the analysis, the topics could be manually

merged for ease of interpretation and improved accuracy. For an example

of such postprocessing, we refer to Supporting Information Appendix F.

Besides the brevity of texts and the high number of (potentially

overlapping) topics, the linguistic diversity is likely a main challenge to

accurate prediction. Nigeria is one of the most linguistically diverse places

in the world (Orekan, 2010), and several different languages are used on

Twitter by stationary users as well as by mobile users. In general, it is to

be expected that code‐switching and code‐mixing on Twitter play a great

role in migration contexts when the social networks of the mobile users

change, for example, when they meet other migrants during their travels

or become acquainted with members of the neighbourhoods they plan to

settle in. Theoretically, our approach is language agnostic, but most

current state‐of‐the‐art models are pretrained on mostly English corpora,

and the predominant language on Twitter itself is English. In contexts

where other languages and different varieties of English are used

(possiblyevenwithinonetweet),thegreater linguistic diversity requires

language models with a higher capacity and flexibility. But besides this

being a technical challenge, there is also an ethical concern. From an

ethical standpoint, inequalities in the representation of languages are

propagated by the dominance of English training data sets: Because

speakers of minority languages are not well represented in the training

data, they are not only less acknowledged, but they are also more likely to

be misunderstood. In the case of African languages, there are efforts

towards better representation in the field of NLP (Masakhane, 2022).

Further, while the use of geospatial SM in approaches such as

ours has the potential to generate substantial benefits by enlarging

the knowledge for policymaking (see Section 5.2), we must

acknowledge it could also be misused in the same fashion.

The data we acquired inform about the affinities and behaviours

of human individuals. Such information necessitates ethical consider-

ations of privacy, consent, anonymity and potential harm that could

arise from the study (Kochupillai et al., 2022; Townsend &

Wallace, 2016). This is particularly the case in the migration context,

as migrants can be a particularly vulnerable group (Sîrbu et al., 2021).

Point aggregation to large SU, as we applied in our processing, is a

way to preserve user's geoprivacy and anonymity (Kounadi &

Resch, 2018) and allows for the ethical use of data without the

acquisition of consent (Williams et al., 2017). No group was smaller

than five users, and SU with less than five detected resident users

were excluded to preserve their anonymity. Consequently, we

believe that privacy and anonymity of users are not at risk by the

presented approach. If follow‐up studies more closely inspect users

and tweets with qualitative methods, it is vital and feasible that the

identity of the users remains undisclosed or cannot be easily

retraced. Also, if future developments allow for significantly higher

spatial precision, geoprivacy should be incorporated into the study

design (see Kounadi & Resch, 2018).

The users' right to be forgotten ought to be respected on a public

platform, such as Twitter, by use of the dedicated batch compliance

API endpoint (Twitter Inc., 2022b). Over the course of this study, we

continuously updated the data and the results, but the present paper

can, of course, not be altered in the same fashion. Thus, we only

present aggregated results wherein no individual can be identified.

We believe that in this manner, the requirements of scientific

research and the rights of data subjects are reconciled.

Finally, ethical concerns result from SM data's ease of use. Compared

with traditional surveys, whether in the field or online, SM data can

informaboutpeopleatreducedcostsandengagement.Itisconceivable

that the researchers' understanding of the geographical contexts is

likewise reduced, and that engagement with the studied population

suffers, while algorithmic biases are propagated. Insights from interviews

and field surveys can help identify and mitigate such issues. Thus, we

argue that SM data should not replace traditional sources of information

but go hand in hand with them in a complementary analysis.

6|CONCLUSION

In this paper we studied the feasibility of analyzing migrant interests

based on geolocated SM data. We found that the data are suitable for the

study of urban demographics and at large spatial and long timescales.

MAST ET AL.

13 of 17

In this context, it allows migration researchers to observe—at a

hitherto unprecedented scale—the subjective perspectives of people

who move, and of those who do not. Flexible in application, our

approach can contribute to migration research by filling a gap

between qualitative studies and large‐scale demographic data, and

support the study of migration as a spectrum rather than a condition.

In our case study we found that for most topics, users' interest is

related to their mobility. It is very likely that the underlying causalities

are complex and can be only understood by integrating the data with

the knowledge of locals and subject‐matter experts. In that light, the

joint analysis of geolocation and texts holds a potential that we have

barely scratched the surface of. We are looking forward to what

future studies in migration research will be able to unearth.

The tweet IDs to reproduce the research can be requested from the

author. The full data are not made available due to ethical restrictions

and the Twitter Developer Agreement and Policy.

ACKNOWLEDGEMENTS

This study has been conducted as part of the European Union's

Caroline Herschel Framework Partnership Agreement on Copernicus

User Uptake (FPCUP) under Grant Agreement No. FPA 275/G/GRO/

COPE/17/10042, Action 2019‐2‐11 “Downstream service/applica-

tion development for monitoring of environmental indicators”, the

project MIGRAWARE, Grant No. 01LG2082C, funded by the German

Federal Ministry of Education and Research (BMBF) as part of the

programme WASCAL WRAP 2.0, and the research and teaching

initiative “A New Focus in English Linguistics: Geolingual Studies”,

Grant No. 98 662, funded by Volkswagen Foundation as part of the

programme “Momentum—Funding for Recently Tenured Professors”.

Open Access funding enabled and organized by Projekt DEAL.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

DATA AVAILABILITY STATEMENT

Thedatathatsupportthefindingsofthisstudyareavailableonrequest

from the corresponding author. The data are not publicly available due to

privacy or ethical restrictions. The implementation of the migration‐

detection algorithm was published as an openly available R‐package

MigrationDetectR (https://github.com/dlr-eoc/ukis-migrationdetectr).

ORCID

Johannes Mast http://orcid.org/0000-0001-6595-5834

Marta Sapena https://orcid.org/0000-0003-3283-319X

Martin Mühlbauer https://orcid.org/0000-0003-3849-1143

Carolin Biewer https://orcid.org/0000-0002-3797-1586

Hannes Taubenböck https://orcid.org/0000-0003-4360-9126

REFERENCES

Afolabi, O. J., Oluwaji, O. A., & Fashola, O. K. (2017). Socio‐economic

impact of road traffic congestion on urban mobility: A case study of

Ikeja Local Government Area of Lagos State, Nigeria. Pacific Journal

of Science and Technology,18(2), 246–255.

Akanle, O., Fayehun, O., & Oyelakin, S. (2021). The information

communication technology, social media, international migration

and migrants' relations with Kin in Nigeria. Journal of Asian and

African Studies,56(6), 1212–1225. https://doi.org/10.1177/

0021909620960148

Armstrong, C., Poorthuis, A., Zook, M., Ruths, D., & Soehl, T. (2021).

Challenges when identifying migration from geo‐located Twitter

data. EPJ Data Science,10(1), 1. https://doi.org/10.1140/epjds/

s13688-020-00254-7

Bastian, M., Hayes, M., Vaughan, W., Shah, S., Skomoroch, P., Kim, H.,

Uryasev, S., & Lloyd, C. (2014). LinkedIn skills: Large‐scale topic

extraction and inference. In Proceedings of the 8th ACM Conference

on Recommender Systems (pp. 1–8). Association for Computing

Machinery (ACM). https://doi.org/10.1145/2645710.2645729

Bloch, R., Fox, S., Monroy, J., & Ojo, A. (2015). Urbanisation and urban

expansion in Nigeria (Vol. 73). ICF International.

Blumenstock, J. E. (2012). Inferring patterns of internal migration from

mobile phone call records: Evidence from Rwanda. Information

Technology for Development,18(2), 107–125. https://doi.org/10.

1080/02681102.2011.643209

Bosco, C., Grubanov‐Boskovic, S., Iacus, S. M., Minora, U., Sermi, F., &

Spyratos, S., (2022). Data innovation in demography, migration and

human mobility (EUR 30907 EN). Publications Office of the European

Union, Luxembourg. https://doi.org/10.2760/027157

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P.,

Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S.,

Herbert‐Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A.,

Ziegler, D., Wu, J., Winter, C., & Amodei, D. (2020). Language models

are few‐shot learners. Advances in Neural Information Processing

Systems,33, 1877–1901.

Calderón, C. A., de la Vega, G., & Herrero, D. B. (2020). Topic modeling

and characterization of hate speech against immigrants on Twitter

around the emergence of a far‐right party in Spain. Social Sciences,

9(11), 188. https://doi.org/10.3390/socsci9110188

Central Intelligence Agency. (2022, June 14). Nigeria—The world factbook.

https://www.cia.gov/the-world-factbook/countries/nigeria/

Charmaz, K. (2006). Constructing grounded theory. Sage Publications.

Chi, G., Lin, F., Chi, G., & Blumenstock, J. (2020). A general approach to

detecting migration events in digital trace data. PLoS ONE,15(10),

e0239408. https://doi.org/10.1371/journal.pone.0239408

Cleveland, W. S., & Loader, C. (1996). Smoothing by local regression:

Principles and methods. In: W. Härdle & M. G. Schimek (Eds.),

Statistical theory and computational aspects of smoothing (pp. 10–49).

Physica HD. https://doi.org/10.1007/978-3-642-48425-4_2

Dekker, R., & Engbersen, G. (2014). How social media transform migrant

networks and facilitate migration. Global Networks,14(4), 401–418.

https://doi.org/10.1111/glob.12040

Devlin, J., Chang, M.‐W., Lee, K., & Toutanova, K. (2018). BERT: Pre‐

training of deep bidirectional transformers for language under-

standing. In Proceedings of the NAACL‐HLT 2019 (pp. 4171–4186).

Association for Computational Linguistics. https://doi.org/10.

48550/arXiv.1810.04805

Didia, D., & Tahir, S. (2022). Enhancing economic growth and government

revenue generation in Nigeria: The role of diaspora remittances. The

Review of Black Political Economy,49(2), 175–202. https://doi.org/

10.1177/00346446211025647

Economic Intelligence Unit, M. of E. P., & B. (2013). The socio‐economic

costs of traffic congestion in Lagos [Working Paper Series, 2].

Efthimion, P. G., Payne, S., & Proferes, N. (2018). Supervised machine

learning bot detection techniques to identify social Twitter bots.

SMU Data Science Review,1(2), 71.

Egghe, L., & Leydesdorff, L. (2009). The relation between Pearson's

correlation coefficient r and Salton's cosine measure. Journal of the

American Society for Information Science and Technology,60(5),

1027–1036.

14 of 17

MAST ET AL.

Fiallos, A., & Jimenes, K. (2019). Using Reddit data for multi‐label text

classification of Twitter users interests. In 2019 Sixth International

Conference on EDemocracy & EGovernment (ICEDEG) (pp. 324–327).

Institute for Electrical and Electronics Engineers (IEEE). https://doi.

org/10.1109/ICEDEG.2019.8734365

Fiorio, L., Abel, G., Cai, J., Zagheni, E., Weber, I., & Vinué, G. (2017). Using

Twitter data to estimate the relationship between short‐term

mobility and long‐term migration. In Proceedings of the 2017 ACM

on Web Science Conference (pp. 103–110). Association for Comput-

ing Machinery (ACM). https://doi.org/10.1145/3091478.3091496

Forenbacher, I., Husnjak, S., Cvitić, I., & Jovović, I. (2019). Determinants of

mobile phone ownership in Nigeria. Telecommunications Policy,43(7),

101812. https://doi.org/10.1016/j.telpol.2019.03.001

Fussell, E., Hunter, L. M., & Gray, C. L. (2014). Measuring the

environmental dimensions of human migration: The demographer's

toolkit. Global Environmental Change,28, 182–191. https://doi.org/

10.1016/j.gloenvcha.2014.07.001

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H.,

Laviolette, F., Marchand, M., & Lempitsky, V. (2016). Domain‐

adversarial training of neural networks. Journal of Machine Learning

Research,17(59), 1–35.

Giachanou, A., & Crestani, F. (2016). Like it or not: A survey of Twitter

sentiment analysis methods. ACM Computing Surveys,49(2), 1–41.

https://doi.org/10.1145/2938640

Gollin, D., Blanchard, P., & Kirchberger, M. (2021). Perpetual motion:

Human mobility and spatial frictions in three African countries (SSRN

Scholarly Paper No. 3960245). https://papers.ssrn.com/abstract=

3960245

Hawelka, B., Sitko, I., Beinat, E., Sobolevsky, S., Kazakopoulos, P., &

Ratti, C. (2014). Geo‐located Twitter as proxy for global mobility

patterns. Cartography and Geographic Information Science,41(3),

260–271. https://doi.org/10.1080/15230406.2014.890072

Heidenreich, T., Eberl, J.‐M., Lind, F., & Boomgaarden, H. (2020). Political

migration discourses on social media: A comparative perspective on

visibility and sentiment across political Facebook accounts in

Europe. Journal of Ethnic and Migration Studies,46(7), 1261–1280.

https://doi.org/10.1080/1369183X.2019.1665990

Hoffmann, R., Šedová, B., & Vinke, K. (2021). Improving the evidence

base: A methodological review of the quantitative climate migration

literature. Global Environmental Change,71, 102367. https://doi.org/

10.1016/j.gloenvcha.2021.102367

Huang, B., & Carley, K. M. (2019). A large‐scale empirical study of

geotagging behavior on Twitter. In Proceedings of the 2019 IEEE/

ACM International Conference on Advances in Social Networks Analysis

and Mining (pp. 365–373). Association for Computing Machinery

(ACM). https://doi.org/10.1145/3341161.3342870

Hübl, F., Cvetojevic, S., Hochmair, H., & Paulus, G. (2017). Analyzing

refugee migration patterns using geo‐tagged tweets. ISPRS

International Journal of Geo‐Information,6(10), 302. https://doi.org/

10.3390/ijgi6100302

Hughes, C., Zagheni, E., Abel, G. J., Sorichetta, A., Wi'sniowski, A.,

Weber, I., & Tatem, A. J. (2016). Inferring migrations: Traditional

methods and new approaches based on mobile phone, social media, and

other big data: Feasibility study on inferring (labour) mobility and

migration in the European union from big data and social media data.

European Commission. https://doi.org/10.2767/61617

Igwe, P. A., Obatolu, A. D. A., Nwajiuba, C. A., Egbo, O. P.,

Ogunnaike, O. O., & Nwekpa, K. C. (2021). The glocalisation of

sports: A study of the influence of European Football Leagues on

Nigerian society. European Journal of International Management,

15(2–3), 247–265. https://doi.org/10.1504/EJIM.2021.113244

International Organization for Migration (IOM) & Displacement Tracking

Matrix (DTM). (2022). Nigeria flood map overview (as of October

2022).https://displacement.iom.int/

Johnson, I. L., Sengupta, S., Schöning, J., & Hecht, B. (2016). The

geography and importance of localness in geotagged social media. In

Proceedings of the 2016 CHI Conference on Human Factors in

Computing Systems (pp. 515–526). Association for Computing

Machinery (ACM). https://doi.org/10.1145/2858036.2858122

Kayastha, T., Gupta, P., & Bhattacharyya, P. (2021). BERT based adverse

drug effect tweet classification. In M. Arjun, K. Ari, M.-E. Antonio, A.

A. Mohammed, A. Ilseyar, M. Zulfat, F.‐M. Eulalia, L. L. Salvador, F.

Ivan, O. Karen, W. Davy, T. Elena, S. Abeed, M. B. Juan, K. Martin, &

G.‐H. Graciela (Eds.), Proceedings of the Sixth Social Media Mining for

Health (#SMM4H) Workshop and Shared Task (pp. 88–90). Associa-

tion for Computational Linguistics. https://doi.org/10.18653/v1/

2021.smm4h-1.15

Kemp, S. (2015). Digital 2015 Nigeria. Datareportal. https://datareportal.

com/reports/digital-2015-nigeria

Kemp, S. (2019). Digital 2019 Nigeria. Datareportal. https://datareportal.

com/reports/digital-2019-nigeria

Khatua, A., & Nejdl, W. (2021). Struggle to settle down! Examining the

voices of migrants and refugees on Twitter platform. In Companion

Publication of the 2021 Conference on Computer Supported Coopera-

tive Work and Social Computing (pp. 95–98). Association for

Computing Machinery (ACM). https://doi.org/10.1145/3462204.

3481773

Kim, J., Sîrbu, A., Giannotti, F., & Gabrielli, L. (2020). Digital footprints of

international migration on Twitter. InM.R.Berthold,Ad.Feelders,&G.

Krempl (Eds.), Advances in intelligent data analysis XVIII (pp. 274–286).

Springer. https://doi.org/10.1007/978-3-030-44584-3_22

Kim, J., Sîrbu, A., Giannotti, F., & Rossetti, G. (2021). Characterising

different communities of Twitter users: Migrants and natives. In

International Conference on Complex Networks and Their Applications.

http://arxiv.org/abs/2103.03710

Kim, J., Sîrbu, A., Giannotti, F., Rossetti, G., & Rapoport, H. (2022). Origin

and destination attachment: Study of cultural integration on Twitter.

EPJ Data Science,11(1), 55.

Kirchberger, M. (2021). Measuring internal migration. Regional Science and

Urban Economics,91, 103714. https://doi.org/10.1016/j.

regsciurbeco.2021.103714

Kirwin, M., & Anderson, J. (2018). Identifying the factors driving West

African Migration [West African Papers No. 17; West African Papers,

Vol. 17]. OECD. https://doi.org/10.1787/eb3b2806-en

Kochupillai, M., Kahl, M., Schmitt, M., Taubenböck, H., & Zhu, X. X. (2022).

Earth observation and artificial intelligence: Understanding emerging

ethical issues and opportunities. IEEE Geoscience and Remote Sensing

Magazine,10(4), 90–124.

Kounadi, O., & Resch, B. (2018). A geoprivacy by design guideline for

research campaigns that use participatory sensing data. Journal of

Empirical Research on Human Research Ethics,13(3), 203–222.

https://doi.org/10.1177/1556264618759877

Kruspe, A., Häberle, M., Hoffmann, E. J., Rode‐Hasinger, S., Abdulahhad, K.,

& Zhu, X. X. (2021). Changes in Twitter geolocations: Insights and

suggestionsforfutureusage.InX.Wei,R.Alan,B.Tim,&R.Afshin(Eds.),

Proceedings of the 2021 EMNLP Workshop W‐NUT (pp. 212–221).

Association for Computational Linguistics. https://doi.org/10.48550/

arXiv.2108.12251

Lamanna, F., Lenormand, M., Salas‐Olmedo, M. H., Romanillos, G.,

Gonçalves, B., & Ramasco, J. J. (2018). Immigrant community

integration in world cities. PLoS ONE,13(3), e0191612. https://doi.

org/10.1371/journal.pone.0191612

Lee, J.‐S., & Nerghes, A. (2018). Refugee or migrant crisis? Labels,

perceived agency, and sentiment polarity in online discussions. Social

Media + Society,4(3), 205630511878563. https://doi.org/10.1177/

2056305118785638

Leydesdorff, L. (2005). Similarity measures, author cocitation analysis, and

information theory. Journal of the American Society for Information

MAST ET AL.

15 of 17

Science and Technology,56(7), 769–772. https://doi.org/10.1002/

asi.20130

Li, L., Goodchild, M. F., & Xu, B. (2013). Spatial, temporal, and

socioeconomic patterns in the use of Twitter and Flickr.

Cartography and Geographic Information Science,40(2), 61–77.

https://doi.org/10.1080/15230406.2013.777139

Malik, M., Lamba, H., Nakos, C., & Pfeffer, J. (2015). Population bias in

geotagged tweets. Proceedings of the International AAAI Conference

on Web and Social Media,9(4), 18–27. https://doi.org/10.1609/

icwsm.v9i4.14688

Masakhane, ∀(2022). Masakhane. Masakhane. https://www.

masakhane.io/

Mau, S., Gülzau, F., Laube, L., & Zaun, N. (2015). The global mobility divide:

How visa policies have evolved over time. Journal of Ethnic and

Migration Studies,41(8), 1192–1213. https://doi.org/10.1080/

1369183X.2015.1005007

Mazzoli, M., Diechtiareff, B., Tugores, A., Wives, W., Adler, N., Colet, P., &

Ramasco, J. J. (2020). Migrant mobility flows characterized with

digital data. PLoS ONE,15(3), e0230264. https://doi.org/10.1371/

journal.pone.0230264

McAuliffe, M., Kitimbo, A., Goossens, A. M., & Ullah, A. A. (2018). World

Migration Report 2018—Chapter 7—Understanding migration journeys

from migrants' perspectives (World Migration Report, 2018(1)).

McAuliffe, M., & Ruhs, M. (2018). World Migration Report 2018—

Introduction (World Migration Report, 1).

McCreadie, R., Buntain, C. L., & Soboroff, I. (2019). TREC incident streams:

Finding actionable information on social media. International Confer-

ence on Information Systems for Crisis Response and Management

(ISCRAM 2019), Valencia, Spain, 19‐22 May 2019. pp. 691–705.

Nguyen, D. Q., Vu, T., & Tuan Nguyen, A. (2020). BERTweet: A pre‐trained

language model for English Tweets. In L. Qun, & S. David (Eds.),

Proceedings of the 2020 Conference on Empirical Methods in Natural

Language Processing: System Demonstrations (pp. 9–14). Association

for Computational Linguistics. https://doi.org/10.18653/v1/2020.

emnlp-demos.2

NiDCOM. (2021). About NiDCOM –Nigerians in diaspora commission.

About NiDCOM. https://nidcom.gov.ng/about-nidcom/

NOI Polls Ltd. (2015). US and UK top list of countries most Nigerians abroad

reside in.https://noi-polls.com/us-and-uk-top-list-of-countries-

most-nigerians-abroad-reside-in-key-reason-for-migration-is-for-

economic-opportunities/

Ojedokun, U. A., Ogunleye, Y. O., & Aderinto, A. A. (2021). Mass

mobilization for police accountability: The case of Nigeria's #End-

SARS protest. Policing: A Journal of Policy and Practice,15(3),

1894–1903. https://doi.org/10.1093/police/paab001

Olteanu, A., Castillo, C., Diaz, F., & Kıcıman, E. (2019). Social data: Biases,

methodological pitfalls, and ethical boundaries. Frontiers in Big Data,

2, 13. https://doi.org/10.3389/fdata.2019.00013

Orabi, M., Mouheb, D., Al Aghbari, Z., & Kamel, I. (2020). Detection of bots in

social media: A systematic review. Information Processing & Management,

57(4), 102250. https://doi.org/10.1016/j.ipm.2020.102250

Orekan, G. (2010). Language policy and educational development in

Africa: The case of Nigeria. Scottish Languages Review,21,17–26.

Pappalardo, L., Simini, F., Rinzivillo, S., Pedreschi, D., Giannotti, F., &

Barabási, A.‐L. (2015). Returners and explorers dichotomy in human

mobility. Nature Communications,6(1), 8166. https://doi.org/10.

1038/ncomms9166

Petutschnig, A., Havas, C. R., Resch, B., Krieger, V., & Ferner, C. (2020).

Exploratory spatiotemporal language analysis of geo‐social network

data for identifying movements of refugees. GI_Forum,1, 137–152.

https://doi.org/10.1553/giscience2020_01_s137

Princewill, N., & Busari, S., CNN. (2021, June). Nigeria bans Twitter after

company deletes president Buhari's tweet. CNN. https://www.cnn.

com/2021/06/04/africa/nigeria-suspends-Twitter-operations-intl/

index.html

Quercia, D., Askham, H., & Crowcroft, J. (2012). TweetLDA: Supervised topic

classification and link prediction in Twitter. In Proceedings of the 3rd

Annual ACM Web Science Conference on—WebSci '12 (pp. 247–250).

Association for Computing Machinery (ACM). https://doi.org/10.1145/

2380718.2380750

Rampazzo, F., Bijak, J., Vitali, A., Weber, I., & Zagheni, E. (2021). A

framework for estimating migrant stocks using digital traces and

survey data: An application in the United Kingdom. Demography,

58(6), 2193–2218. https://doi.org/10.1215/00703370-9578562

Ramponi, A., & Plank, B. (2020). Neural unsupervised domain adaptation

in NLP—A survey. In S. Donia, B. Nuria, & Z. Chengqing (Eds.),

Proceedings of the 28th International Conference on Computational

Linguistics (pp. 6838–6855). International Committee on Computational

Linguistics. https://doi.org/10.18653/v1/2020.coling-main.603

Reips, U.‐D., & Buffardi, L. E. (2012). Studying migrants with the help of

the Internet: Methods from psychology. Journal of Ethnic and

Migration Studies,38(9), 1405–1424. https://doi.org/10.1080/

1369183X.2012.698208

Rowe, F., Mahony, M., Graells‐Garrido, E., Rango, M., & Sievers, N. (2021).

Using Twitter to track immigration sentiment during early stages of

the COVID‐19 pandemic. Data & Policy,3, e36. https://doi.org/10.

1017/dap.2021.38

Schewel, K. (2020). Understanding immobility: Moving beyond the

mobility bias in migration studies. International Migration Review,

54(2), 328–355. https://doi.org/10.1177/0197918319831952

Sîrbu, A., Andrienko, G., Andrienko, N., Boldrini, C., Conti, M., Giannotti, F.,

Guidotti, R., Bertoli, S., Kim, J., Muntean, C. I., Pappalardo, L.,

Passarella, A., Pedreschi, D., Pollacci, L., Pratesi, F., & Sharma, R.

(2021). Human migration: The big data perspective. International

Journal of Data Science and Analytics,11(4), 341–360. https://doi.

org/10.1007/s41060-020-00213-5

Spyratos, S., Vespe, M., Natale, F., Weber, I., Zagheni, E., & Rango, M.

(2018). Migration data using social media: A European perspective.

Publications Office of the European Union. https://hdl.handle.net/

21.11116/0000-0004-7C55-2

Subrahmanian, V. S., Azaria, A., Durst, S., Kagan, V., Galstyan, A.,

Lerman, K., Zhu, L., Ferrara, E., Flammini, A., & Menczer, F. (2016).

The DARPA Twitter bot challenge. Computer,49(6), 38–46. https://

doi.org/10.1109/MC.2016.183

Taubenböck, H., Staab, J., Zhu, X. X., Geiß, C., Dech, S., & Wurm, M.

(2018). Are the poor digitally left behind? Analyzing urban divides

using remote sensing and Twitter data. ISPRS International Journal of

Geo‐Information,7(8), 304.

Torch Contributors. (2019). CrossEntropyLoss—PyTorch 1.11.0 documenta-

tion.https://pytorch.org/docs/stable/generated/torch.nn.

CrossEntropyLoss.html

Townsend, L., & Wallace, C. (2016). Social media research: A guide to ethics

(pp. 1–16). University of Aberdeen.

Twitter Inc. (2022a). Twitter API for academic research | products.https://

developer.Twitter.com/en/products/Twitter-api/academic-research

Twitter Inc. (2022b, June). Batch compliance.https://developer.Twitter.

com/en/docs/Twitter-api/compliance/batch-compliance/

introduction

United Nations Department of Economic and Social Affairs. (1998).

Recommendations on statistics of international migration. Revision 1.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N.,

Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. Advances

in Neural Information Processing Systems,30, 5998–6008. https://doi.

org/10.48550/arXiv.1706.03762

Wang, Z., Hale, S. A., Adelani, D., Grabowicz, P. A., Hartmann, T., Flöck, F.,

& Jurgens, D. (2019). Demographic inference and representative

population estimates from multilingual social media data. In The

World Wide Web Conference (pp. 2056–2067). Association for

Computing Machinery (ACM). https://doi.org/10.1145/3308558.

3313684

16 of 17

MAST ET AL.

Willekens, F. (2019). Evidence‐based monitoring of international migra-

tion flows in Europe. Journal of Official Statistics,35(1), 231–277.

https://doi.org/10.2478/jos-2019-0011

Willekens, F., Massey, D., Raymer, J., & Beauchemin, C. (2016).

International migration under the microscope. Science,352(6288),

897–899. https://doi.org/10.1126/science.aaf6545

Williams, M. L., Burnap, P., & Sloan, L. (2017). Towards an ethical framework

for publishing Twitter data in social research: Taking into account users'

views, online context and algorithmic estimation. Sociology,51(6),

1149–1168. https://doi.org/10.1177/003803851770814

World Bank. (2022). Remittances brave global headwinds. Special focus: Climate

migration.Knomad.https://www.knomad.org/sites/default/files/

publication-doc/migration_and_development_brief_37_nov_2022.pdf

Zagheni, E., Garimella, V. R. K., Weber, I., & State, B. (2014). Inferring

international and internal migration patterns from Twitter data. In

Proceedings of the 23rd International Conference on World Wide Web

(pp. 439–444). Association for Computing Machinery (ACM).

https://doi.org/10.1145/2567948.2576930

Zhu, X. X., Wang, Y., Kochupillai, M., Werner, M., Haberle, M.,

Hoffmann, E. J., Taubenbock, H., Tuia, D., Levering, A., Jacobs, N.,

Kruspe, A., & Abdulahhad, K. (2022). Geoinformation harvesting

from social media data: A community remote sensing approach. IEEE

Geoscience and Remote Sensing Magazine,10(4), 150–180. https://

doi.org/10.1109/MGRS.2022.3219584

SUPPORTING INFORMATION

Additional supporting information can be found online in the

Supporting Information section at the end of this article.

How to cite this article: Mast, J., Sapena, M., Mühlbauer, M.,

Biewer, C., & Taubenböck, H. (2023). The migrant

perspective: Measuring migrants' movements and interests

using geolocated tweets. Population, Space and Place, e2732.

https://doi.org/10.1002/psp.2732

MAST ET AL.

17 of 17

Geospatiality: the effect of topics on the presence of geolocation in English text data

Article

Feb 2025
INT J GEOGR INF SCI

Geolocated text data are a promising data source for spatial analyses in many fields, from disease surveillance to the spatial humanities. This study investigates the relationship between texts’ thematic categories and their likelihood of containing usable geolocation information by quantifying and modelling this relationship across seven diverse English text datasets of different types, including web forums, microblogs, news, and magazines. We find that the likelihood of geoinformation is highly variant, being high for the category ‘Travel, Tourism & Migration’ and low for ‘Private Life, Family & Relationships’. The rank-correlation of this likelihood between datasets is moderate to strong. These findings indicate that the topic plays a significant role in determining the frequency of geospatial references within the text, and that the effect is not entirely dataset-specific. This contributes to the empirical study of the concept of spatiality and provides valuable insights for bias mitigation in the increasing use of text as data for spatial analyses.

The voices of the displaced: Mobility and Twitter conversations of migrants of Ukraine in 2022

Article

Full-text available

Jan 2024
INFORM PROCESS MANAG

Monitoring and understanding human migration as triggered by a crisis is challenging. Combining spatial analysis with natural language processing when analyzing social media data helps to understand the mobility and the needs of migrants better. For this paper, we used geo-located Twitter data to analyze the mobility of and topics discussed by migrants of the Ukraine war in 2022. We removed bots, accounts showing implausible mobility, and automated text content from our dataset. Then, we applied a transformer-based multilingual topic modeling framework to identify the migrants' discourses. We assessed the topics discussed by migrants before leaving Ukraine, after leaving Ukraine and after returning to Ukraine. Our results show that "Attack reports", "politics", "donations to Ukrainians", "food export/production", "human-itarian aid", "nuclear threat", "Ukrainian places", "job search", and "war journalism" were dominant topics before leaving from and after returning to Ukraine. "Food", "social media", "transport", "art", and "finance", however, were important topics right after leaving the country. Overall, our results reveal plausible spatial patterns of migration, which are similar to those reported by official statistics (R 2 = 0.89), showing the reliability of geotagged social media data to monitor human mobility. This information can complement official sources, adding first-hand information on the mobility and needs of migrants across space, time, topics, and languages. This is crucial to develop humanitarian response plans when time is of the essence.

Origin and destination attachment: study of cultural integration on Twitter

Article

Full-text available

Nov 2022

The cultural integration of immigrants conditions their overall socio-economic integration as well as natives’ attitudes towards globalisation in general and immigration in particular. At the same time, excessive integration—or assimilation—can be detrimental in that it implies forfeiting one’s ties to the origin country and eventually translates into a loss of diversity (from the viewpoint of host countries) and of global connections (from the viewpoint of both host and home countries). Cultural integration can be described using two dimensions: the preservation of links to the origin country and culture, which we call origin attachment , and the creation of new links together with the adoption of cultural traits from the new residence country, which we call destination attachment . In this paper we introduce a means to quantify these two aspects based on Twitter data. We build origin and destination attachment indices and analyse their possible determinants (e.g., language proximity, distance between countries), also in relation to Hofstede’s cultural dimension scores. The results stress the importance of language: a common language between origin and destination countries favours origin attachment, as does low proficiency in the host language. Common geographical borders seem to favour both origin and destination attachment. Regarding cultural dimensions, larger differences among origin and destination countries in terms of Individualism, Masculinity and Uncertainty appear to favour destination attachment and lower origin attachment.

Earth Observation and Artificial Intelligence: Understanding Emerging Ethical Issues and Opportunities

Article

Full-text available

Oct 2022

Ethics is a central and growing concern in all applications utilizing Artificial Intelligence (AI). Earth Observation (EO) or Remote Sensing (RS) research relies heavily on both Big Data and AI or Machine Learning (ML). While this reliance is not new, with increasing image resolutions and the growing number of EO/RS use-cases that have a direct impact on governance, policy, and the lives of people, ethical issues are taking center stage. In this article, we provide scientists engaged with AI4EO research (i) a practically useful overview of the key ethical issues emerging in this field with concrete examples from within EO/RS to explain these issues, and (ii) a first road-map (flowchart) that scientists can use to identify ethical issues in their ongoing research. With this, we aim to sensitize scientists about these issues and create a bridge to facilitate constructive and regular communication between scientists engaged in AI4EO research on the one hand, and ethics research on the other. The article also provides detailed illustrations from four AI4EO research fields to explain how scientists can redesign research questions to more effectively grab ethical opportunities to address real-world problems that are otherwise akin to ethical dilemmas with no win-win solution in sight. The paper concludes by providing recommendations to institutions that want to support ethically mindful AI4EO research and provides suggestions for future research in this field.

Using Twitter to track immigration sentiment during early stages of the COVID-19 pandemic

Article

Full-text available

Dec 2021

Large-scale coordinated efforts have been dedicated to understanding the global health and economic implications of the COVID-19 pandemic. Yet, the rapid spread of discrimination and xenophobia against specific populations has largely been neglected. Understanding public attitudes toward migration is essential to counter discrimination against immigrants and promote social cohesion. Traditional data sources to monitor public opinion are often limited, notably due to slow collection and release activities. New forms of data, particularly from social media, can help overcome these limitations. While some bias exists, social media data are produced at an unprecedented temporal frequency, geographical granularity, are collected globally and accessible in real-time. Drawing on a data set of 30.39 million tweets and natural language processing, this article aims to measure shifts in public sentiment opinion about migration during early stages of the COVID-19 pandemic in Germany, Italy, Spain, the United Kingdom, and the United States. Results show an increase of migration-related Tweets along with COVID-19 cases during national lockdowns in all five countries. Yet, we found no evidence of a significant increase in anti-immigration sentiment, as rises in the volume of negative messages are offset by comparable increases in positive messages. Additionally, we presented evidence of growing social polarization concerning migration, showing high concentrations of strongly positive and strongly negative sentiments.

Changes in Twitter geolocations: Insights and suggestions for future usage

Conference Paper

Full-text available

Nov 2021

Twitter data has become established as a valuable source of data for various application scenarios in the past years. For many such applications, it is necessary to know where Twitter posts (tweets) were sent from or what location they refer to. Researchers have frequently used exact coordinates provided in a small percentage of tweets, but Twitter removed the option to share these coordinates in mid-2019. Moreover, there is reason to suspect that a large share of the provided coordinates did not correspond to GPS coordinates of the user even before that. In this paper, we explain the situation and the 2019 policy change and shed light on the various options of still obtaining location information from tweets. We provide usage statistics including changes over time, and analyze what the removal of exact coordinates means for various common research tasks performed with Twitter data. Finally, we make suggestions for future research requiring geolocated tweets.

Geoinformation Harvesting From Social Media Data: A community remote sensing approach

Article

Dec 2022

As unconventional sources of geoinformation, massive imagery and text messages from open platforms and social media form a temporally quasi-seamless, spatially multiperspective stream, but with unknown and diverse quality. Due to its complementarity to remote sensing (RS) data, geoinformation from these sources offers promising perspectives, but harvesting is not trivial due to its data characteristics. In this article, we address key aspects in the field, including data availability, analysis-ready data preparation and data management, geoinformation extraction from social media text messages and images, and the fusion of social media and RS data. We then showcase some exemplary geographic applications. In addition, we present the first extensive discussion of ethical considerations of social media data in the context of geoinformation harvesting and geographic applications. With this effort, we wish to stimulate curiosity and lay the groundwork for researchers who intend to explore social media data for geoapplications. We encourage the community to join forces by sharing their code and data.

Struggle to Settle down! Examining the Voices of Migrants and Refugees on Twitter Platform

Conference Paper

Oct 2021

Improving the evidence base: A methodological review of the quantitative climate migration literature

Article

Nov 2021
GLOBAL ENVIRON CHANG

The question whether and how climatic factors influence human migration has gained both academic and public interest in the past years. Based on two meta-analyses, this paper systematically reviews the quantitative empirical literature on climate-related migration from a methodological perspective. In total, information from 127 original micro- and macro-level studies is analyzed to assess how different concepts, research designs, and analytical methods shape our understanding of climate migration. We provide an overview of common methodological approaches and present evidence on their potential implications for the estimation of climatic impacts. We identify five key challenges, which relate to the i) measurement of migration and ii) climatic events, iii) the integration and aggregation of data, iv) the identification of causal relationships, and v) the exploration of contextual influences and mechanisms. Advances in research and modelling are discussed together with best practice cases to provide guidance to researchers studying the climate-migration nexus. We recommend for future empirical studies to employ approaches that are of relevance for and reflect local contexts, ensuring high levels of comparability and transparency.

Measuring internal migration

Article

Jul 2021
REG SCI URBAN ECON

Martina Kirchberger

Measuring the extent and direction of the movement of people across space is key to understanding spatial gaps and processes of structural transformation. A range of sources from census data, survey data, to administrative data and newer types of data such as mobile phone data have been used to examine such patterns. This paper investigates how different types of data are used to measure migration and mobility, and how complementary data determine which research questions they can answer. I discuss measurement concerns related to these different types of data, definitions of migration rates used in the literature, selection concerns and other survey design considerations. I conclude by highlighting gaps in the measurement of migration.

Enhancing Economic Growth and Government Revenue Generation in Nigeria: The Role of Diaspora Remittances

Article

Jun 2021

Even though remittances constitute the second-largest source of foreign exchange for Nigeria, with a $24 billion inflow in 2018, its impact on economic growth remains unclear. This study, therefore, examined the short-run and long-run impact of remittances on the economic growth of Nigeria using the vector error correction model. Utilizing World Bank data covering 1990–2018, the empirical analysis revealed that remittances hurt economic growth in the short run while having no impact on economic growth in the long run. Our parameter estimates indicate that a 1% increase in remittances would result in a 0.9% decrease in the gross domestic product growth rate in the short run. One policy implication of this study is that Nigeria needs to devise policies and interventions that minimize the emigration of skilled professionals rather than depending on remittances that do not offset the losses to the economy due to brain drain.

BERT based Adverse Drug Effect Tweet Classification

Conference Paper

Jan 2021

The migrant perspective: Measuring migrants' movements and interests using geolocated tweets

Abstract and Figures

Recommended publications

My knowledge is shaping the future.

DLR research in and from space

DLR research for the energy system of the future

Get to know the German Aerospace Center as an Employer

The voices of the displaced: Mobility and Twitter conversations of migrants of Ukraine in 2022

Enhancing Earth Observation of Migration with Insights from Social Media

Geospatiality: the effect of topics on the presence of geolocation in English text data

Can Social Media Data Help to Understand the Socio-spatial Heterogeneity of the Interests and Concer...