Evaluating socio-economic state of a country analyzing airtime credit and mobile phone datasets

Transcription

1 1 Evluting socio-economic stte of country nlyzing irtime credit nd moile phone dtsets Thorlf Gutierrez Gutier Krings Vincent D. Blondel ICTEAM Rel Impct Anlytics ICTEAM Université ctholique de Louvin Université ctholique de Louvin rxiv: v1 [cs.cy] 17 Sep 2013 Astrct Relile sttisticl informtion is importnt to mke politicl decisions on sound sis nd to help mesure the impct of policies. Unfortuntely, sttistics offices in developing countries hve scrce resources nd sttisticl censuses re therefore conducted spordiclly. Bsed on moile phone communictions nd history of irtime credit purchses, we estimte the reltive income of individuls, the diversity nd inequlity of income, nd n indictor for socioeconomic segregtion for finegrined regions of n Africn country. Our study shows how to use moile phone dtsets s strting point to understnd the socio-economic stte of country, which cn e especilly useful in countries with few resources to conduct lrge surveys. Index Terms moile phones, ICTs for development, socioeconomic level, welth distriution, poverty mp, communities, socio-economic segregtion. I. INTRODUCTION Sttisticl informtion is importnt to foster wreness of the socil, demogrphic, economic nd environmentl conditions of country. They help forge opinions, mke politicl nd evidence-sed decisions on sound foundtions. They re lso useful to help llocte resources. Unfortuntely, mny developing countries do not hve much up-to-dte sttisticl informtion out the stte of their popultion. Surveys re onerous, they re not conducted very often. For exmple, the lst popultion census in Angol ws conducted in 1970 [1]. Even when they re conducted, their reliility is questionle (see Appendix). Indeed, the figures of officil reports usully involve gret del of guesswork [2], s individuls nd enterprises re less likely to e officilly registered nd to keep forml records of their economic ctivities. Up-to-dte reports re of primry importnce in Côte d Ivoire, s the recent civil wr hs chnged the fce of the country. With uiquitous sensors like moile phones, we hve the opportunity to get fct-sed figures t the individul level, nd ggregte them to get locl informtion cross the country. Furthermore, the logs of moile phone ctivity re generted in rel-time nd cn therefore e used s n erly lrm system, insted of witing for the next survey. As these logs re lredy eing generted nd stored for legl purposes, they re virtully free compred to censuses, nd could thus e of gret use to developing countries tht hve few resources. In this work, we explore moile phone logs nd irtime credit purchses to identify nd extrct proxies for welth in Côte d Ivoire. Our study is sed on 2012 logs from lrge moile opertor in Côte d Ivoire tht hve client se representtive of the popultion. The dtset contins the Cll Detil Records for clls nd text messges rising from clients. They hve timestmp, n ID for the inititor, n ID for the receiver nd n identifier for the cell tower used. The dtset lso contins the timestmp nd mount of every irtime credit purchse mde y every client. Firstly, we summrize results ville in the literture on similr questions. Secondly, we explore the irtime credit purchses of users nd oserve ehviorl ptterns in order to identify proxy for welth. We then mp the verge of the proxy with its vrition nd distriution inequlity. Finlly, we look t the dt from socil networking ngle. We detect the underlying communities in the network nd oserve degree of welth homophili for the memers of the sme community. II. RELATED WORK As moile phone dtsets hve een incresingly mde ville to reserch tems, links etween moile phone usge nd socio-economic levels hve een studied y mny. Egle et l. hve developed new metric to cpture the socil diversity of communiction ties within n individul s socil network [3], where high diversity scores imply tht n individul splits his time more evenly mong socil ties. The uthors showed tht, in the United Kingdom, socil diversity scores were strongly correlted with the government s Index of Multiple Deprivtion rnk. They showed the sme result for the numer of structurl holes [4] spnned y n indivudl s links. Wng et l. proposed mesure of dydic reciprocity tht cptures the degree of communictive imlnce etween two nodes [5]. The index is miniml when two nodes hve exctly the sme proility of clling ech other, nd increses monotoniclly with the difference etween proilities. Reserchers t UN Glol Pulse explined

2 Percentge of users 2 how this reciprocity index hd een linked to higher socioeconomic levels [13]. Blumenstock et l. [6] inferred household expenditures of smll, ut representtive, set of users from phone survey in Rwnd. They then nlyzed the correltions with moile phone usge nd household expenditures of the queried users. They found correltions, minly with the numer of interntionl clls, nd the numer of different districts contcted. They lso found positive correltion with the verge irtime credit purchse, ut they insist tht they did not use roust pproch nd hope to do so in future work. Blumenstock & Egle [7] did similr nlysis, ut compred the usge of the lower nd upper qurtiles of people, sed on their estimte of socio-economic level. They showed tht richer people hd higher mximum vlue of irtime credit purchses thn poorer people. They did not give numers for the verge of irtime credit purchses. Soto et l. [8] nd Fris-Mrtinez et l. [9] hve focused on moility vriles. They looked t mny of them, nd reported the following top fetures : the numer of cell tower used per week, the mximum distnce trveled in week nd the weekly rdius of gyrtion (i.e. the typicl rnge of user trjectory [10]). Both were le to predict socio-economic clsses of neighorhoods in their city of study. Smith et l. [12] nd Mo et l. [11] lso worked on Côte d Ivoire. Smith et l. found strong correltion etween census dt on poverty nd the following vriles : grvity residuls or the sum of differences etween oserved nd expected flows using grvity model, the numer of clls nd totl durtion, the socil diversity s defined y Egle et l., nd introversion or the rtio of self-flow to totl flow. Mo et l. did similr nlysis, ut sed on different census dt. They oserved the highest correltion with socio-economic levels using the rtio of weighted out-degree to the sum of weighted out nd in-degree. Unfortuntely, we rgue in the Appendix tht the census dt from Côte d Ivoire, on which the uthors sed their models, re questionle. A hint of this is given y the difference etween their poverty mps. Furthermore, the censuses re t lest four yers older thn the uthors dtsets, nd were conducted efore the civil wr. III. PURCHASES We strt y nlyzing irtime credit purchses. We first look t ech individul s ehvior, y quntifying the vrition in purchse mounts of ech user with their Coefficient of Vrition (CV). It is equl to the rtio of the stndrd devition σ to the verge µ. CV = σ/µ This vriility mesure is independent of the series verge, nd trnsltes into the fct tht it would e s esy or difficult for someone to mke purchse twice the size of wht he is 100% 80% 60% 40% 20% 0% 0% 20% 40% 60% 80% 100% Coefficient of Vrition of user's pruchses Fig. 1: Cumultive frequency nlysis of Coefficients of Vrition of ech user s purchses used to, whtever the mount he is used to. Figure 1 shows cumultive frequency nlysis of CVs. We see tht 80% of the users hve CV under 62%, nd 97% of the users hve one under 100%. As reference, the CV of the series of ll purchses of every client is 238%. The mounts purchsed y most people pper to e stle. To simplify the dt, we thus ssign n verge purchse per user, without losing much informtion. We oserve different types of users. Some people mke few ig purchses, nd others mke mny smll purchses. Our hypothesis is tht this difference in ehvior predicts household income. The users who mke ig purchses will e richer thn the users who mke multiple smll ones, simply ecuse the poorer will not hve enough csh flow to uy lots of irtime credit ll t once. Although there hs not yet een pulished study tht vlidtes this hypothesis, some studies hve shown correltion etween mounts of purchses nd socio-economic levels [6], [7]. Also, crriers hve confirmed tht, through mrketing surveys they hve developed relile predictive models for household income y looking t size nd frequency of credit purchses [13]. If we hd some up-to-dte nd relile dt out welth or household income in the country, we could lso uild predictive model of welth sed on size nd frequency of purchses. But, s we explin in the Appendix, we do not hve such dt. Without trining set, we do not hve ny insight into how to comine or weigh the size nd frequency. We cn therefore only look t one vrile t time. Between the size nd frequency of purchses, we choose to look t the size of purchses, s it reflects the most user s income or welth. Someone who is poor will hve to uy irtime credit in smll mounts while someone who is rich cn mke lrger purchses. We thus mp the purchse verges in Figure 2. Aidjn hs higher verge thn most of the country, which ws expected with its sttus of economic cpitl of Côte d Ivoire nd the iggest se port in West Afric. On the other hnd, the other cities do not hve purchse verge tht stnds out. The Lierin order in the South-West is unexpectedly welthy. Indeed, the popultion density is low nd there is no pprent reson for the region to e welthier

3 3 c d Averge of Purchse verges (CFA) Urn res Coefficient of Vrition of Purchse verges (%) Urn res Fig. 2: Averge of ech user s purchse verge. () Aidjn, () Lierin order (c) to Mli nd Burkin Fso (d) Rod to Ghn thn nother rurl re. On the contrry, it is known for its insecurity nd lnd conflicts [14]. However, some of the welth could come from illegl ctivities, s it seems to e loction of choice for drug, rms nd humn trfficking [15], [16]. The orders with the min rods to Mli nd Burkin Fso in the North nd to Accr (Ghn) in the South-Est lso stnd out. As they re ig routes of pssge, it is resonle to expect tht these plces would stimulte economic ctivity. Generlly the cost in the South is lso welthier, which could e explined y ctivities linked to tourism. These hypothesis re only specultion : we do not know the cuslity of these results nd it is difficult to interpret them with little informtion out the context. Nonetheless, this mp could e useful s n indictor of potentilly interesting further investigtions. Furthermore, this mp hs to e nlyzed with cre, s low popultion density res re more sensitive to outliers. For exmple, some plces seem welthy, ut y looking closer t the unpopulted ntenn, it is plced on ig qurry with dozen of exptrite or executives tht completely skew the verge Fig. 3: Coefficient of Vrition of purchse verges, mesures the diversity. () Aidjn, () Divo (c) Korhogo We cn mp our estimte of welth diversity nd inequlity with the Coefficient of Vrition nd the Gini index of purchse verges, respectively in Figures 3 nd 4. Most urn res clerly stnd out in terms of diversity nd inequlity, confirming tht they host poor nd rich people. Korhogo, on the other hnd, is city s ig s the others [17] ut does not hve s much diversity or inequlity s them : people tht mke smll purchses mke up much igger proportion of its popultion thn in other cities. Divo is lmost the sme, ut its wider mid-clss gives it higher CV. The Lierin order does not hve ny diversity, indicting tht ll the people living there do uy purchses in lrger mounts. IV. S OCIAL NETWORK The Cll Detil Records tell us who communictes with whom. With this informtion, we cn uild the socil network of users. Ech node in the network corresponds to client, nd two clients re connected if they communicted with ech other t lest once per month. We do not impose reciprocity, i.e. tht there re communictions going in oth directions. Indeed, we preferred omitting this condition, s imposing it resulted in communities with memers living further from ech other. We prefer when they live closer to ech other, s the center of grvity of communities will then represent the position of its memers more ccurtely, nd our mpping will lso e more ccurte.

4 4 e Gini index of Purchse verges Urn res Urn(res Weigthed(verge(of Coefficient(of(Vrition((u) within(communities Fig. 4: Gini index of purchse verges, mesures inequlity. () Aidjn, () Divo (c) Korhogo Not enforcing reciprocity hs some implictions, the noise is not s filtered nd, therefore, some links do not represent socil interction. The min outliers re hotlines, helpdesks nd service numers which end up in the socil network with degrees ove the thousnds. We identify them with their disproportionte difference etween in nd out-degree, or when they contcted too mny distinct people. Regrding links, they re weighted ccording to the sum of the numer of clls nd text messges exchnged etween the two users. The resulting network is highly clustered despite its sprseness. Indeed, its verge locl clustering coefficient, s defined in [18], is equl to On the other hnd, the shuffled network (i.e. the network with the sme degree distriution, ut with links redistriuted rndomly) presented n verge locl clustering coefficient of only The links within the socil network re thus not scttered rndomly, nd we cn detect the communities in the network using the Louvin method [19] with scled modulrity [20] to detect smll group of friends. Compring the different memers of communities, we find certin homophili in terms of purchse verges. Consider the Coefficients of Vrition of purchse verges within ech community, the weighted verge of these vlues with the size of ech community is equl to 48.5%. If we distriute the Fig. 5: Weighted verge of Coefficient of Vrition of verge purchses within ech community. () Aidjn, () Bouke (c) Sn Pe dro (d) Dlo (e) Ymoussoukro purchse verges rndomly throughout the socil network, we find weighted verge of 68.5%. This confirms tht people tend to e friends with people tht hve the sme purchse verge s themselves. In Figure 5, we mp these vlues to see which plces hve more diverse communities thn others. Some cities tend to hve more diverse communities, notly Aidjn, Bouke nd Sn Pe dro. This mkes them cities where poor individuls would wnt to go to, s these re the plces where they hve the most proility of eing prt of community of welthier people nd could, t the sme time, clim the socio-economic ldder. Other cities hve high diversity (see Figure 3), without hving diversity in their communities, giving us hint of plces where socio-economic segregtion is stronger thn others. Among these cities re Dlo nd Ymoussoukro. V. C ONCLUSION We nlyzed irtime credit purchsing ehviors, nd noted tht most individuls uy credit y mounts of similr sizes. We hypothesize tht size nd frequency of purchses re correlted with the income of individuls. But, in sence of relile dt to uild model tht tkes oth vriles into ccount, we used the verge size of purchses to estimte the

5 5 reltive welth of n individul. As expected, the economic cpitl clerly stood out in terms of welth. Other plces tht stood out were the cost, with Sn Pédro, the orders with min rods to Mli, Burkin Fso nd Ghn, nd ll the Lierin order. Other cities do not look welthier thn other rurl res. In terms of diversity, urn res clerly stood out, with some of them hving more inequlity thn others in the distriution of our proxy for welth. We detected communities in the socil network, nd found tht people in the sme community tend to hve the sme purchse verge. Indeed, the verge Coefficient of Vrition of purchse verge within communities is equl to 48.5%, compred to 68.5% when we shuffle the purchse verges. We showed wy to extrct indictors of welth, inequlity nd segregtion out of moile phone dtset, nd we think tht there is rel opportunity to e seized in terms of otining low-cost socio-economic sttistics from the uiquitous sensors tht re moile phones. These logs of moile phone opertors cn indeed e used to uild predictive models, ut privte or officil (ut relile) surveys still need to e conducted to help uild more ccurte models nd ssess their qulity. Nevertheless, these dtsets cn e strting point to understnd the socio-economic stte of the popultion in countries with no resources to conduct lrge surveys, thus helping orgniztions to mke etter decisions founded on fct-sed sttistics. REFERENCES [1] Ttem, A., nd C. Linrd. (2011). Popultion mpping of poor countries. Nture 474, no : [2] Jerven, M. (2013). Poor Numers: How We re Misled y Africn Development Sttistics nd wht to Do out it. Cornell University Press. [3] Egle, N., M. Mcy, nd R. Clxton. (2010). Network diversity nd economic development. Science 328, no. 5981: [4] Burt, R. S. (2009). Structurl holes: The socil structure of competition. Hrvrd University Press. [5] Wng, C., O. Lizrdo, D. Hchen, A. Strthmn, Z. A. Toroczki, nd N. V. Chwl. (2013). A dydic reciprocity index for repeted interction networks. Network Science, 1, pp doi: /nws [6] Blumenstock, J. E., Y. Shen, nd N. Egle. (2010). A method for estimting the reltionship etween phone use nd welth. In QulMeetsQunt Workshop t the 4th Interntionl Conference on Informtion nd Communiction Technologies nd Development. [7] Blumenstock, J. E., nd N. Egle. (2012). Divided we cll: Disprities in ccess nd use of moile phones in Rwnd. Informtion Technologies & Interntionl Development 8, no. 2 : pp-1. [8] Soto, V., V. Fris-Mrtinez, J. Virsed, nd E. Fris-Mrtinez. (2011). Prediction of socioeconomic levels using cell phone records. In User Modeling, Adption nd Personliztion, pp Springer Berlin Heidelerg, [9] Fris-Mrtinez, V., J. Virsed, nd E. Fris-Mrtinez. (2010). Socioeconomic levels nd humn moility. In Qul Meets Qunt Workshop- QMQ. [10] Gonzlez, M. C., C. A. Hidlgo, nd A.-L. Brsi. (2008). Understnding individul humn moility ptterns. Nture 453, no : [11] Mo, H., X. Shui, Y.-Y. Ahn, nd J. Bollen. (2013). Moile Communictions Revel the Regionl Economy in Côte d Ivoire. Presented t NetMo 2013 for the D4D Chllenge. [12] Smith, C., A. Mshhdi, nd L. Cpr. (2013). Uiquitous Sensing for Mpping Poverty in Developing Countries. Presented t NetMo 2013 for the D4D Chllenge. [13] Kirkptrick, R. (2012). Beyond Trgeted Ads: Big Dt for Better World. Tlk t O Reilly Strt Conference [14] Norwegin Refugee Council. (2012) Conflits fonciers et sécurité limentire dns l région frontlière entre le Liéri et l Côte d Ivoire. ISBN [15] United Ntions Office on Drugs nd Crime. Lieri [Internet]. Pulished etween 2005 nd [Cited 2013 My 15]. Aville from http: //it.ly/181rzvn [16] U.S. Deprtment of Stte. Trfficking in Persons Report : Office to monitor nd comt trfficking in persons [Internet]. 4 Jun 2008 [Cited 2013 My 15]. Aville from [17] Institut Ntionl de l Sttistique (Répulique de Côte d Ivoire). Côte d Ivoire - Recensement générl de l popultion et de l hittion 1998 RGPH 98. [18] Soffer, S. N., nd A. Vázquez. (2005). Network clustering coefficient without degree-correltion ises. Physicl Review E 71, no. 5 : [19] Blondel, V. D., J.-L. Guillume, R. Lmiotte, nd E. Lefevre. (2008). Fst unfolding of communities in lrge networks. Journl of Sttisticl Mechnics: Theory nd Experiment 2008, no. 10 (2008): P [20] Lmiotte, Renud, J-C. Delvenne, nd Muricio Brhon. (2008). Lplcin dynmics nd multiscle modulr structure in networks. rxiv preprint rxiv: APPENDIX UNRELIABILITY OF CENSUS REPORTS We would hve liked to compre our dt with officil reports sed on income census like the ones performed y the Ntionl Sttistics Office of Côte d Ivoire in 2002 nd The poverty rtes of these two reports re given in Figure 6. Unfortuntely, we find tht either the sitution hs completely chnged in six yers time or the estimted poverty rtes re unrelile. If the poverty rtes re unrelile, then it is difficult to conclude nything from the comprison. If the sitution hs chnged so much in six yers, it is very likely tht it hs chnged lot in the four yers time seprting the census nd our dtset, even more so with the civil wr in etween. In this cse lso, it is difficult to conclude nything from the comprison.

6 Fig. 6: Poverty rtes in 2002 nd 2008 of Côte d Ivoire, ccording to censuses from the Ntionl Sttistics Office of Côte d Ivoire. (Relistion : Institut Ntionl de l Sttistique (INS), Division crtogrphie) 6