Estimating individual employment status using mobile phone network data

Abstract

This study provides the first confirmation that individual employment status can be predicted from standard mobile phone network logs externally validated with household survey data. Individual welfare and households vulnerability to shocks are intimately connected to employment status and professions of household breadwinners. At a societal level unemployment is an important indicator of the performance of an economy. By deriving a broad set of novel mobile phone network indicators reflecting users financial, social and mobility patterns we show how machine learning models can be used to predict 18 categories of profession in a South-Asian developing country. The model predicts individual unemployment status with 70.4 percent accuracy. We further show how unemployment can be aggregated from individual level and mapped geographically at cell tower resolution, providing a promising approach to map labor market economic indicators, and the distribution of economic productivity and vulnerability between censuses, especially in heterogeneous urban areas. The method also provides a promising approach to support data collection on vulnerable populations, which are frequently under-represented in official surveys.

Full text

Estimating individual employment status
using mobile phone network data
1Pål Sundsøy, 1Johannes Bjelland, 1Bjørn-Atle Reme, 2Eaman Jahani, 3,4Erik Wetter, 3Linus Bengtsson
1Telenor Group Research, 2MIT Media Lab, 3Flowminder Foundation, 4Stockholm School of Economics
ABSTRACT
This study provides the first confirmation that individual
employment status can be predicted from standard mobile phone
network logs externally validated with household survey data.
Individual welfare and households’ vulnerability to shocks are
intimately connected to employment status and professions of
household breadwinners. At a societal level unemployment is an
important indicator of the performance of an economy. By
deriving a broad set of novel mobile phone network indicators
reflecting users’ financial, social and mobility patterns we show
how machine learning models can be used to predict 18 categories
of profession in a South-Asian developing country. The model
predicts individual unemployment status with 70.4% accuracy.
We further show how unemployment can be aggregated from
individual level and mapped geographically at cell tower
resolution, providing a promising approach to map labor market
economic indicators, and the distribution of economic
productivity and vulnerability between censuses, especially in
heterogeneous urban areas. The method also provides a promising
approach to support data collection on vulnerable populations,
which are frequently under-represented in official surveys.
Keywords
Big-Data Development, machine learning, unemployment, socio-
economic indicators, mobile phone metadata, profession
1. INTRODUCTION
Unemployment is a key indicator of labor market performance
[1,2]. When workers are unemployed, their families also get
affected, while the nation as a whole loses their contribution to the
economy in terms of the goods and services that could have been
produced [3]. Unemployed workers also lose their purchasing
power, which can lead to the unemployment for other workers,
creating a cascading effect that ripples through the economy [4].
Additionally, unemployment has been shown to be a driver of
interregional migration patterns [5].
Counting each and every unemployed person on a monthly basis
would be a very expensive, time-consuming and impractical
exercise. In many countries, such as US, a monthly population
survey is run to measure the extent of unemployment in the nation
[6]. In developing countries such surveys often tend to have a low
spatial and temporal frequency [7]. Lacking statistics may lead to
higher uncertainties in economic outlook, lower purchasing
capacity and higher burden of debt. The problems of
unemployment and poverty have always been major obstacles to
economic development [8], and proper background statistics is
important to change this trend.
The increasing availability and reliability of new data sources, and
the growing demand of comprehensive, up-to-date international
employment data are therefore of high priority. Specifically,
privately held data sources have been shown to hold great promise
and opportunity for economic research, due to both high spatial
and temporal granularity [9].
One of the most promising rich new data sources is mobile phone
network logs [10], which have the potential to deliver near real-
time information of human behavior on individual and societal
scale [11]. The prediction from mobile phone metadata are vast
given that more than half of the world’s population now own a
mobile phone. Several research studies have used large-scale
mobile phone metadata, in the form of call detail records (CDR)
and airtime purchases (top-up) to quantify various socio-economic
dimensions. On aggregated level mobile phone data have shown
to provide proxy indicators for assessing regional poverty levels
[12,13], illiteracy [14], population estimates [15], human
migration [16,17] and epidemic spreading [18]. On individual
level mobile phone data have been used to predict, among others,
socio-economic status [19,20], demographics [21,22] and
personality [23].
Two previous papers analyze employment trends through cell
phone data. The work by [24] argue that unemployment rates may
be predicted two-to-eight weeks prior to the release of traditional
estimates and predict future rates up to four months ahead of
official reports accurately than using historical data alone. The
other study [25] shows that mobile phone indicators are associated
with unemployment and this relationship is robust when
controlling for district area, population and mobile penetration
rate. The results of these analyses highlight the importance of
investigating the relationship between mobile phone data and
employment data further.
Our work separates from [24,25] in several ways:
(1) Bottom-up approach: we focus on predicting employment
status on the individual level to be able get a clean view of
the main drivers and ability to discover global predictors that
are useful across various employment groups. This is of
relevance also as previous research has uncovered non-linear
relationships between worker flows and job flows at the
micro level, indicating a more complex relationship between
the micro and macro levels of employment statistics than
simple aggregation [26].
(2) Geography: We focus on a low HDI South-Asian
developing country where employment statistics is almost
non-existing and highly in demand.
(3) Method: Individually matched mobile phone data and large-
scale survey data allow us to carefully optimize machine-
learning models for multiple professions.
(4) Data: In addition to CDR data we also include airtime
purchases, as financial proxy, for our analysis.
The rest of this paper is organized as follows: In section 2 we
describe the methodological approach, including the features and
modeling approach. In section 3 we describe the results. In

chapter 4 we discuss the limitations from a holistic perspective,
while we finally draw our conclusions.
2. APPROACH
2.1 Data
Household survey data: Data from two nationally representative
cross-sectional household surveys of 200,000 individuals in a
low-income South Asian country was analyzed. The data was
collected for business intelligence purposes at time Q114 and
Q214 by an external survey company commissioned by the
operator. The survey discriminated between 18 types of
professions for the head of household, including currently being
retired and unemployed. The head of household’s was asked for
his or her most frequently used phone number. 87% of households
in the country has at least one mobile phone.
Mobile phone data: Mobile phone logs for 76 000 of the
surveyed 200 000 individuals belonging to the leading operator
were retrieved from a period of six months and de-identified by
the operator. Individual level features were built from the raw
mobile phone data and was subsequently coupled with the
corresponding de-identified phone numbers from the survey. The
social features were subsetted from a graph consisting of in total
113 million subscribers and 2.7 billion social ties. No content of
messages or calls were accessible and all individual level data
remained with the operator.
The following sub-sections describe the features and the machine
learning algorithm used for our prediction.
2.2 Features
Independent variables: The independent variables are built
entirely from the cell phone datasets. A structured dataset
consisting of 160 novel mobile phone features are built from the
raw CDRs and airtime purchases, and categorized into three
dimensions: (1) financial (2) mobility and (3) social features
(Table 1). The features are customized to be predictive of
employment status and include various parameters of the
corresponding distributions such as weekly or monthly median,
mean and variance. In addition to basic features such as incoming
and outgoing MMS, voice, SMS, internet and video calls we
investigate more customized features such as the consumption rate
of airtime purchases (spending speed), the amount time spent on
each base station, the size of social circle, the time spent on
different contacts and features related to the phone type owned by
the customer.
Dependent variables: The dependent variables were built from
the survey data. Since our aim is to separate one specific
profession/employment status from the others we build 18 binary
classifiers – one for each profession (student/non-student,
employed/unemployed and so forth). These classifiers are then
trained separately.
Table 1. Sample of independent features from mobile phone
metadata used in model
Dimension
Features
Financial
Airtime purchases: Recharge amount per
transaction, Spending speed, fraction of
lowest/highest recharge amount, coefficient of
variation recharge amount etc
Revenue: Charge of outgoing/incoming SMS,
MMS, voice, video, value added sevices,
roaming, internet etc.
Handset: Manufacturer, brand, camera enabled,
smart/feature/basic phone etc
Mobility
Home district/tower, radius of gyration, entropy
of places, number of places visited etc.
Social
Social Network: Interaction per contact, degree,
entropy of contacts etc.
General phone usage: Out/In voice duration,
SMS count, Internet volume/count, MMS count,
video count/duration, value-added services
duration/count etc.
2.3 Deep Learning Models
We tested our features against several algorithms such as gradient
boosted machines (GBM), random forest (RF), support vector
machines (SVM) and K-nearest neighbors (kNN). Based on the
performance of individual algorithms we propose a standard
multi-layer feedforward neural network architecture where the
weighted combination of the n input signals is aggregated, and an
output signal f(α) is transmitted by the connected neuron. The
function f used for the nonlinear activation is rectifier f(α) ≈
log(1+eα ) . To minimize the loss function we apply a standard
stochastic gradient descent with the gradient computed via back-
propagation. We use dropout as a regularization technique to
prevent over-fitting [27]. Dropout secures that each training
example is used in a different model which all share the same
global parameters. For the input layer we use the value of 0.1 and
0.2 for the hidden layers. In total 18 models are built – one for
each pre-classified profession type.
To compensate class imbalance, the minority class in the training
set is up-sampled. The minority class is then randomly sampled,
with replacement, to be of the same size as the majority class. In
our set-up, each model is trained and tested using a 75%/25%
(training/testing) split. Commonly used performance metrics for
classification problems [28], including overall accuracy,
sensitivity and specificity, are reported for the test-set.

3. RESULTS
3.1 Individual employment status
The average prediction accuracy for all 18 profession groups were
67.5%, with clerk being the easiest to predict with accuracy of
73.5%, and skilled worker the most difficult (accuracy: 61.9%)
(Fig. 1a).
Our unemployment model predicted whether phone users were
unemployed with an accuracy of 70.4% (95% CI:70.1-70.6%).
The accuracy difference between the training and test-set was
3.6%, which indicate our trained model has good generalization
power. The true positive rate (sensitivity/recall) was 67% and true
negative rate (specificity) 70.4%. Given the original baseline of
2.1%, the model predicts unemployment on average 30 times
better than random.
Each cross-validated model was subsequently restricted to use its
20 most important predictors (applying Gedeon method [29]). An
investigation of the five most important predictors for each
profession is given in Fig 2b. This network shows how the
professions are linked together via common predictors. We
observe that several features are predictive across multiple
professions – indicated by high in-degree in the network.
Predictors that are superior across multiple professions include the
most frequently used cell tower (longitude and latitude): in the
case of unemployed (red node) this signal indicate that the model
may catch regions of low economic development status, e.g. slum
areas where unemployment is high. Other cross-profession
predictors include number of visited places, the radius of gyration
(how far the person usually travels from his home tower) and
recharge amount per transaction. These indicators have earlier
been shown to be important financial proxy indicators for
household income in underdeveloped Asian markets [20].
Unemployed people tend further to have few interactions with
their friends, generate more voice calls at night (when calls are
(A) (B)
Figure 1: A) Test-set performance of predicting individual profession from mobile phone data. For each profession accuracy ,
sensitivity and specificity are reported. Top behavioral predictors are indicated with colors (excluding most used tower (lon,lat) )
B) Network of professions linked via common predictors. Each profession links to their 5 most important predictors. The link width
is proportional to the scaled relative importance of the current predictor. For example the most important predictors for being
unemployed (highlighted in red node color) are most used cell tower (lon,lat), interaction per contact, nocturnal voice (%) and number
of voice calls.

cheaper) and make less voice calls. They also tend to top-up with
the lowest recharge amount per transaction (sixth predictor) - a
feature that also occurs as a predictor of low household income
[20].
As seen in the figure, students have the most unique predictive
signal, when it comes to few overlapping predictors with other
groups. Interestingly they are not the easiest group to predict,
which indicate the Deep Learning models have found stronger
non-linear relationships in other categories of profession.
3.2 Geographical employment mapping
A natural next step is to move from individual employment to
geographical distribution of employment rates and professions.
Large Asian cities are typically covered by thousands of mobile
phone towers opening up to the possibility of providing a detailed
spatial understanding of differences in employment rates and
profession types. In figure 2 we have mapped out the predicted
geographical distribution of employment rate per home tower, in
one of the large cities for six different groups. The individual
employment rates are here calculated using the test set, aggregated
Figure 2 Geographical distributions of employment status and profession categories per base station in one of the larger Asian cities
with over 1,500 cell towers and 18 million people. Employment rates are calculated by using the out-of-sample test set, aggregated and
averaged to their respective home tower. Individual prediction accuracy and top three predictors are given to the right for A) Unemployed
B) Teachers and students C) Landlords D) Retired E) Clerks.

and averaged to the tower from which most calls were made
between 7pm and 5am (defined as home tower). Fig 2a shows that
the unemployed population are spatially spread out across the city,
indicating many pockets of unemployment that traditional surveys
would not easily pick up. The retired population (Fig 2d) and
landlords (Fig 2c) are spatially more concentrated.
By investigating the top 3 most important behavioral predictors
we interestingly observe that the physical width of handset is the
most important for the retired people. The elderly in this market
tends to have older and smaller phones than the younger
generation. This in contrast to Scandinavian markets where larger
handsets are marketed especially towards elderly people. We also
notice that SMS and mobile Internet consumption are the best
indicators for being a student (Fig 2b). The most important
predictor of Clerks (Fig 2e) is a low radius of gyration (or
mobility radius – reflecting static office jobs). They also tend to
have more advanced consumption patterns (more variation in top-
up refill types) and longer voice duration. We also notice that
airtime purchase information is among top predictors for several
professions, stressing the importance to include such datasets in
future research.
4. DISCUSSION
This study shows how employment status can be predicted from
mobile phone logs, purely by investigating users’ metadata. By
deriving economic, social and mobility features for each mobile
user we predict individual employment status with up to 73.5%
accuracy. We address how various profession groups relates in a
network via industry standard mobile network indicators, and
further show how individual employment can be aggregated and
mapped geographically with high spatial resolution on cell tower
level. The geospatial indicators available in mobile network data
additionally allow promising avenues for research on two major
topics related to the economic impacts of employment status; the
increased productivity effects that have been found to be an
outcome of increased spatial density of employment [30], and the
interregional migration patterns that are related to unemployment
[5].
One general concern in such studies is always the sampling
selection bias. A large data set may make the sampling rate
irrelevant, but it doesn’t necessarily make it representative. The
fact that the people who use mobile phone are not necessarily a
representative sample of the larger population considered. This
issue is especially of high relevance when considering how
mobile phone data may be used for monitoring, economic
forecasting and development. Research studies are often based
purely on data from one mobile operator, and depending on the
type of data one can expect individuals to be represented
disproportionally with respect to certain characteristics. These
problems persist even if data from all operators in a country were
available, nearing the total population. In our study we consider
data from one large operator, where customers have to own a
mobile phone to be counted. We argue, however, that there should
be a high correlation between population employment and sample
employment, since our sample is large, and most people in the
country own a mobile phone (more than 85% penetration rate).
Additional sources for external validation were not easily
obtainable.
An important policy application of this work is the prediction of
regional and individual employment rates in developing countries
where official statistics is limited or non-existing.
5. REFERENCES
1.
Lovati, J.: The unemployment rate as an economic indicator.
Federal reserve bank of st.louis (1976)
2.
Keynes, Maynard: The General Theory of Employment,
Interest and Money. Basingstoke, Hampshire: Palgrave
Macmillan. ISBN 0-230-00476-8. (2009)
3.
International Labour Organization: Global Unemployment
Trends. (2013)
4.
Garegnani: Heterogeneous Capital, the Production Function
and the Theory of Distribution. Review of Economic Studies.
37 (3): 407–436 (1970)
5.
Pissarides, C., Wadsworth, J.: The Flow Approach to Labor
Markets: New Data Sources and Micro-Macro links. The
Journal of Economic Perspectives, Vol. 20, No. 3, pp. 3-
26(24) (1989)
6.
U.S. Bureau of Labor Statistics: How the Government
Measures Unemployment. (2014)
7.
ILO: World Employment Social Outlook. (2015)
8.
Economy Watch: Unemployment and Poverty.
http://www.economywatch.com/unemployment/poverty.html
(2010)
9.
Einav, L., Levin, J.: Economics in the age of big data.
Science, Vol 346(6210) DOI: 10.1126/science.1243089
(2014)
10.
Lokanathan, Gunaratne: Behavioral insights for development
from Mobile Network Big Data: enlightening policy makers
on the State of the Art. In : Available at:
http://dx.doi.org/10.2139/ssrn.2522814 (2014)
11.
Lazer, Pentland, Adamic, Aral, Barabasi, Brewer, Christakis:
Computational Social Science. Science. Vol. 323, Issue 5915,
pp. 721-723 (2009)
12.
Blumenstock, Cadamuro: Predicting poverty and wealth from
mobile phone metadata. Science VOL 350 ISSUE 6264 1
(2015)
13.
Steele, J.: Predicting poverty using cell phone and satellite
data. In submission (2016)
14.
Sundsøy, P.: Can Mobile Usage predict illiteracy in a
developing country? arXiv Preprint: 1607.01337 (2016)
15.
Deville, Linard, Martin, Gilbert, Stevens, Gaughan, Blondel,
Tatem: Dynamic population mapping using mobile phone
data. PNAS 15888–15893, doi: 10.1073/pnas.1408439111
(2014)
16.
Lu, X.: Detecting climate adaptation with mobile network
data in Bangladesh: anomalies in communication, mobility
and consumption patterns during cyclone Mahasen. Climatic
Change Volume 138, Issue 3, pp 505–519 (2016)
17.
Lu, X.: Unveiling hidden migration and mobility patterns in
climate stressed regions: A longitudinal study of six million
anonymous mobile phone users in Bangladesh. Global
Environmental Change Volume 38, May 2016, Pages 1–7
(2016)

18.
Wesolowski, Qureshi, Boni, Sundsøy, Johansson, Rasheed,
Engø-Monsen, Buckee: Impact of human mobility on the
emergence of dengue epidemics in Pakistan. PNAS 11887-
11892 DOI: 10.1073/PNAS.1504964112 (2015)
19.
Guitierrrez, Krings, Blondel: Evaluating socio-economic state
of a country analyzing airtime credit and mobile phone
datasets. arXiv preprint arXiv:1309.4496 (2013)
20.
Sundsøy, P., Bjelland, J., Reme, B., Iqbal, A., Jahani, E.:
Deep learning applied to mobile phone data for Individual
income classification. ICAITA doi:10.2991/icaita-16.2016.24
(2016)
21.
Felbo, B., Pentland, S., Sundsøy, P., Montjoye, Y., Lehmann,
S.: Using Deep Learning to predict demographics from mobile
phone metadata. arXiv:1511.06660v4 (2016)
22.
Jahani, Sundsøy, Bjelland, Pentland, Montjoye: Improving
official statistics in Emerging Markets using Machine
Learning and Mobile phone data. In submission (2016)
23.
Montjoye, Y.-A., Quoidbach, Robic, Pentland: Predicting
Personality Using Novel Mobile Phone-Based Metrics.
Volume 7812 of the series Lecture Notes in Computer Science
pp 48-55 (2013)
24.
Toole, J., Lin, Y.-r., Muehlegger, E., Shoag, D., Gonzalez, M.,
Lazer, D.: Tracking employment shocks using mobile phone
data. Journal of the royal society interface. DOI:
10.1098/rsif.2015.0185 (2015)
25.
Almatouq, Castrillo, Pentland: Mobile Communication
Signatures of Unemployment. arXiv:1609.01778 [cs.SI]
(2016)
26.
Faberman, D., Haltiwanger, J.: The Flow Approach to Labor
Markets: New Data Sources and Micro-Macro links. The
Journal of Economic Perspectives, Vol. 20, No. 3, pp. 3-
26(24) (2006)
27.
Dahl, G.: Improving Deep Neural Networks for LVCSR using
Rectified Linear Units and Dropout. ICASSP, 8609-8613.
(2013)
28.
Koyejo, O.: Consistent Binary Classification with Generalized
Performance Metrics. NIPS (2014)
29.
Gedeon, T.: Data Mining of inputs: analysing magnitude and
functional measures. International Journal of Neural Systems,
vol 8, no. 2, pp. 209-218. (1997)
30.
Ciccone, A., Hall, R.: Productivity and density of economic
activity. American Economic Review, Vol. 86, No. 1, pp. 54-
70. (1996)
31.
OECD: Main Economic Indicators. (2016)