1 Introduction
Wider penetration of the mobile phone services in the developing countries is providing alternate ways to provide financial services through mobile phones (“Mobile money services”) in these countries. Considering the lower cost of additional infrastructure government, mobile operators and development agencies all around the world have been trying to promote mobile money services. However, only a few countries have seen widespread adoption of mobile money. Most of the moible money services around the world have been unable to achieve a critical user mass. scharwatt_state_2014 . An important question thus revolves around understanding what drives customers to adopt and use mobile money. This paper is focused on the following question: How do patterns of adoption of mobile money vary across different demographics (gender, urban/rural areas; rich/ poor areas)?
We try to answer this question by applying machine learning algorithms over social networks extracted from mobile money transactions.
2 Related Work
The work on the adoption of digital financial services has been dominated by macroeconomic work related to regulatory issues around interoperability and the logistics of mobile money agents dermish_branchless_2011 ; donovan_mobile_2012 and a very few research papers have explored the adoption of mobile money from a quantitative perspective. However, mining of insights from mobile communication meta-data has been a popular area of research and some examples of work in this area include predicting the socioeconomic status blumenstock_predicting_2015 , gender frias2010gender of mobile phone subscribers, customer churn behavior khan2015behavioral and analysis of gender disparities using social networks extracted from mobile communication logs reed2016gender .
We have used deterministic finite automata (DFA) based feature engineering over CDR data which is quite similar to the approach used in khan_dfa_2016 and blumenstock_predicting_2015 . Lastly, the differences in the usage of technology across men and women, poor & rich, and the urban & rural population has been a popular theme of research in the ICTD domain jackson_race_2008 , blumenstock2010mobile ,etc. In comparison to these studies, our work employs a more comprehensive feature engineering approach over a bigger dataset to evaluate the role of the different type of features in the adoption of mobile money for people with different demographic backgrounds.
3 Context and Methods
Our input data consists of anonymized call detail records (CDR) and mobile money transaction information (MMTR) from a major mobile operator of Pakistan. Each row of the CDR typically consists of the tuple containing {callerID, recipientID, date, time, duration, callerLocation, receiverLocation}. We categorize each user either as a “Voice only user",“Registered mobile money user" or “P2P mobile money user” as explained below.
-
Voice only users: The users who do not use mobile money services at all.
-
Registered mobile money users: The users who have signed up for mobile money.
-
P2P mobile money users: The users who have either sent or received money to other users.
Instead of relying on a few handpicked features, we wanted to use a comprehensive set of features for which we have used the DFA based feature generation algorithm as shown in Figure 1. As an example, say we are interested in constructing a feature for each individual
that corresponds to, “the variance in the average duration of outgoing calls made by
on different days of the week”. This feature can be constructed through the following set of recursive rules: (a) filter outgoing calls, (b) filter transactions initiated by , (c) group by day of week, (d) focus on call duration, (e) aggregate by average (duration per day of week), (f) aggregate using variance(over average daily durations). DFA also helps in clear categorization of the features. For the purpose of this work, we use the following categories: usage features, mobility features and network (structural properties of the networks) features. We have performed two different type of supervised learning experiments:
Voice Only Users vs. Registered Mobile Money Users & Voice Only Users vs. Peer-to-peer Mobile Money UsersFor each of these experiments, we drew a stratified random balanced sample of adopters/peer-to-peer users and non-adopters for each of the six categories (Males, Females; Urban, Rural; and Rich, Poor ). Gradient boosting algorithm
friedman2001greedyis used to classify users and determine prominent features.







4 Results
The distribution of conditional feature importance values for all the features generated through the DFA is shown in the Figures 4, 4, 4. First thing to note in the Figures 4, 4, 4 is that different category of features have different aggregate feature importance for different cases. Both network and usage related features are a more likely indicator of adoption for females while mobility related features have more predictive power for male users. Usage related features are the most important category for the female users while the mobility related features are the most important one for male users.
Similarly, mobility is the most important category of features for both urban and rural districts. However, compared to rural districts, the network, and usage related features of the users in the urban districts are not that important. The importance of the network related features for the rural users indicates that the awareness of the mobile money services and endorsement effects from other mobile money users play an important role in the adoption of mobile money services. The trends for the rich vs. poor districts are quite similar to the urban and rural districts.
The features importance for each of the experiments for voice only users vs peer to peer mobile money users is shown in the figures 7, 7 and 7. Usage related features are the top category of peer-to-peer transactions for the female users, while the mobility related features are the top determinant for the male users as shown in the Figure 7. Mobility related features are still the top features for the urban and rural districts (7) while in comparison to the Figure 7 the importance of network related features is lower while the importance of the usage features is higher for the rural district’s users. In comparison to the Figure 4, Figure 7 shows that the mobility related features are the most important one for the users in the rich districts while the usage related features are the most important determinant for peer to peer transactions in the poor districts.


Cross-validated accuracy for predicting the adoption or the peer to peer usage of mobile money obtained through gradient boosting is shown in Figure 9.
Index | Feature | Category |
Panel A: Gender: Female | ||
1 | Variance in the number of active days with incoming SMS | Usage |
2 | Rank Percentile of the Size of the incoming SMS contact network | Network |
Panel A: Gender: Male | ||
1 | Number of unique contact locations on weekdays | Mobility |
2 | Variance in the contact locations per day | Mobility |
Panel C: Rural | ||
1 | Number of unique contact locations on weekend | Mobility |
2 | Percentile of the network size on weekend sms | Network |
Panel D: Urban | ||
1 | Number of unique contact locations on weekdays | Mobility |
2 | Variance in the duration of the weekend calls to each contact locations | Usage |
Panel E: Poor | ||
1 | Number of unique contact locations on week days | Mobility |
2 | Size of the network per contact locations per day | Network |
Panel F: Rich | ||
1 | Number of unique contact locations on the weekdays | Mobility |
2 | Number of unique contact locations on the weekend | Mobility |
-
Notes: Rank percentile for a feature indicate the percentile of the feature value for the subscriber as compared to all the contacts.
5 Conclusion
Most interesting finding of this work are the differences in performance of the different type of features for different sections of the society. For example, the usage category was the top one for the female users but not for the male users. This indicates that the males who may be working remotely or have to commute to work are more likely to use the mobile money, whereas the women who are more active or have higher technology literacy are more likely to user mobile money. Similarly, mobility of the users is the prime determinant when it comes to the adoption of mobile money in both urban & rural and rich & poor districts. However, the people in rural districts with a larger network are more likely to use mobile money. These results can help the marketing and developing agencies to promote the adoption of mobile money in the developing countries.
References
- [1] J. Blumenstock, G. Cadamuro, and R. On. Predicting poverty and wealth from mobile phone metadata. Science, 350(6264):1073–1076, 2015.
- [2] J. Blumenstock and N. Eagle. Mobile divides: gender, socioeconomic status, and mobile phone use in rwanda. In Proceedings of the 4th ACM/IEEE International Conference on Information and Communication Technologies and Development, page 6. ACM, 2010.
- [3] A. Dermish, C. Kneiding, P. Leishman, and I. Mas. Branchless and mobile banking solutions for the poor: A survey of the literature. Innovations: Technology, Governance, Globalization, 6(4):81–98, Oct. 2011.
- [4] K. Donovan. Mobile money for financial inclusion. Information and Communications for Development, pages 61–73. The World Bank, 2012.
- [5] V. Frias-Martinez, E. Frias-Martinez, and N. Oliver. A gender-centric analysis of calling behavior in a developing economy using call detail records. 2010.
- [6] J. H. Friedman. Greedy function approximation: a gradient boosting machine. Annals of statistics, pages 1189–1232, 2001.
- [7] L. A. Jackson, Y. Zhao, A. Kolenic III, H. E. Fitzgerald, R. Harold, and A. Von Eye. Race, gender, and information technology use: the new digital divide. CyberPsychology & Behavior, 11(4):437–442, 2008.
- [8] M. R. Khan and J. E. Blumenstock. Predictors without borders: Behavioral modeling of product adoption in three developing countries. In Proceedings of the 22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York, NY, USA, 2016. ACM.
- [9] M. R. Khan, J. Manoj, A. Singh, and J. Blumenstock. Behavioral modeling for churn prediction: Early indicators and accurate predictors of custom defection and loyalty. In 2015 IEEE International Congress on BigData, pages 677–680. IEEE, 2015.
- [10] P. J. Reed, M. R. Khan, and J. Blumenstock. Observing gender dynamics and disparities with mobile phone metadata. In Proceedings of the Eighth International Conference on Information and Communication Technologies and Development, New York, NY, USA, 2016. ACM.
- [11] C. Scharwatt, A. Katakam, J. Frydrych, A. Murphy, and N. Naghavi. State of the Industry: Mobile Financial Services for the Unbanked. Technical report, GSMA, 2014.