Spatiotemporal Route Estimation Consistent with Human Mobility Using Cellular Network Data

Transcription

1 Internatonal Workshop on the Impact of Human Moblty n Pervasve Systems and Applcatons 203, San Dego (8 March 203) Spatotemporal oute Estmaton onsstent wth Human Moblty Usng ellular Network Data Hrosh Kanasug, Yoshhde Sekmoto, Mor Kurokawa 2, Takafum Watanabe 2, Shgek Muramatsu 2 and yosuke Shbasak The Unversty of Tokyo, hba, Japan 2 KDDI &D Laboratores Inc., Satama, Japan yok@css.u-tokyo.ac.jp, sekmoto@css.u-tokyo.ac.jp, mo-kurokawa@kddlabs.jp, tk-watanabe@kddlabs.jp, mura@kddlabs.jp, shba@css.u-tokyo.ac.jp Abstract ontnuous personal poston nformaton has been attractng attenton n a varety of servce and research areas. In recent years, many studes have appled the telecommuncaton hstores of moble phones (Ds: call detal records) to poston acquston. Although large-scale and long-term data are accumulated from Ds through everyday use of moble phones, the spatal resoluton of Ds s lower than that of exstng postonng technologes. Therefore, nterpolatng spatotemporal postons of such sparse Ds n accordance wth human behavor models wll facltate servces and researches. In ths paper, we propose a new method to compensate for D drawbacks n trackng postons. We generate as many canddate routes as possble n the spatotemporal doman usng trp patterns nterpolated usng road and ralway networks and select the most lkely route from them. Trp patterns are feasble combnatons between stay places that are detected from ndvdual locaton hstores n Ds. The most lkely route could be estmated through comparng canddate routes to observed Ds durng a target day. We also show the assessment of our method usng Ds and GPS logs obtaned n the expermental survey. Keywords-spatotemporal route estmaton; personal moblty; Ds I. INTODUTION ontnuous personal poston nformaton has been attractng attenton n varous servce and research areas. Although embeddng GPS functonalty n moble phones has become a common means of acqurng the postons of ther users, ssues such as power consumpton and ndoor postonng reman. As a soluton to the aforementoned ssues, n recent years, the telecommuncaton hstores of moble phones (Ds: call detal records) have been appled to poston acquston and personal behavor analyss. Ds are recorded n the exstng nfrastructures of moble phone carrers; therefore, they can be used to elmnate the addtonal workload for moble phones durng data acquston and to acqure large-scale and long-term data on all telecommuncaton users. Moble phone carrers should release some Ds even f there are certan restrctons and f agreement to the terms of usage from each moble phone user must be obtaned; nevertheless, the use of Ds s attractve and expected n some research felds relevant to human moblty, such as transportaton, dsaster reducton, and urban development. Orgnally, the purpose of recordng Ds has been to dscover problems wth telecommuncaton nfrastructures and resolve them expedtously. And also Ds have been fundamental for generatng bllng data. When makng a call, sendng a text, or browsng the nternet usng a moble phone, tmestamp and poston nformaton of the connected base staton are recorded as D data. Therefore, the temporal resoluton of Ds dffers for each person accordng to the moble phone communcaton pattern. In addton, the spatal resoluton of Ds s lower than that of exstng postonng devces such as GPS because they depend on the coverage area of the base statons. Accordngly, nterpolatng the spatotemporal postons of such sparse Ds n accordance wth actual human behavor wll facltate relevant servces and researches. In ths study, we attempt to estmate a user s personal route over an entre day based on spatotemporal smlarty usng Ds as spatally sparse personal footprnts. As the basc prncple, the estmaton s conducted by generatng as many canddate routes as possble usng trp patterns nterpolated usng the shortest path of road and ralway networks, and selectng the most lkely route from them. In partcular, canddate routes n the spatal doman are exhaustvely generated based on stay places detected from ndvdual locaton hstores n Ds and trp patterns consstng of the shortest paths between them. The spatal route s then determned by selectng the nearest canddate route based on the shape from the trajectory of Ds on the target day of estmaton. When generatng the canddate route, we only employ the shortest paths and exclude the ndvdual dversty of route selecton and traffc nformaton. Addtonally, to consder the dversty of the occurrence tme of the trps n days, the temporal patterns of the spatal route are generated by shftng the occurrence tme of the trps. Fnally, the most lkely route n the spatotemporal doman could be estmated by comparng the lkelhood values of the observed Ds on the target day. The proposed estmaton method s assessed by Ds and GPS logs obtaned from the expermental survey that s conducted for 25 days wth 84 examnees. The remander of the paper s as follows. In secton II, we descrbe related work based on Ds n partcular, and n /3/$ IEEE 267

2 secton III, we explan the proposed methods. In secton IV, we present the results of the experment. Fnally, n secton V, we summarze the results and suggest future drectons for research. II. ELATED WOKS Personal poston nformaton, f obtaned contnuously and accumulated regonally, helps grasp personal moblty characterstcs and gan an overvew of the tme-varyng populaton dstrbuton. Ashbrook et al. obtaned long-term GPS data on several people and predcted ther moblty patterns by estmatng sgnfcant locatons from GPS data [3]. However, these data dd not nvolve temporal characterstcs or specfc trajectores. Although Froehlch et al. also acqured long-term GPS data from hundreds of partcpants, they attempted to predct drvng routes by clusterng ndvdual trps extracted from GPS logs excludng temporal characterstcs [4]. Because there are numerous exstng studes related to moblty estmaton usng GPS logs, some studes would be applcable to D-based estmaton by regardng Ds as spare poston data. Gven that Ds already cover a wdespread area globally and that telecommuncatons data on ndvduals have been contnuously recorded, many studes have employed Ds for moblty analyss. Some studes have provded overvews of tme-varyng populaton dstrbutons by summarzng and estmatng the number of people stayng around specfc areas [6][0]. On the other hand, by regardng Ds as footprnts of personal movements, other studes have attempted to predct people s moblty by extractng sgnfcant locatons such as the orgns and destnatons of trps [5][7][8][9]. Although most of the studes usually organze Ds nto statstcs around certan areas or base statons, there are few studes attemptng to estmate consstent spatotemporal routes from Ds. III. METHODS In ths secton, as well as through Fgure, we descrbe methods for spatotemporal route estmaton as follows. In subsecton A, we explan the method to detect stay places from Ds, and n subsecton B, we descrbe how canddate routes are generated from stay places and how the most approprate route s dentfed from them. Subsequently, n subsecton, we present the method of temporal pattern estmaton through reallocaton of the duraton of stay. Fnally, the valdaton method for the estmated route wth Fgure. Outlne of route estmaton procedure GPS logs s summarzed n subsecton D. A. Stay Place Detecton from Ds Here, we descrbe methods to dentfy stay places and tme ntervals of each user based on ndvdual locaton hstores {P (t):t = t 0,,t n }, where P (t) represents a twodmensonal spatal pont at tme t n D data. Frst, to nvestgate whether the user stayed or moved wthn each tme segment, we employed a unform approach for tme segmentaton usng sldng wndows wth a wdth T and a shft S. In other words, let the ntal tme be t 0. Then, the frst tme segment s [t 0, t 0 +T), the second tme segment s [t 0 + S, t 0 + T + S), etc. The mean-shft procedure [] s then appled to each pont belongng to each tme segment. The meanshft procedure s used to seek the mode of the densty of spatal ponts. Every teraton of the procedure calculates the mean m(p (t)) of nearby ponts of P (t) wthn a wndow determned by a wndow functon K( ) and shfts ponts P (t) to m(p (t)) untl they converge. P ( t') K( P ( t') P ( t)) P ( t) m ( P ( t)) = () K ( P ( t') P ( t)) P ( t') As for the wndow functon K( ), we use a rectangular wndow defned as follows: K(P (t)) = f x < h, and 0 otherwse, where the parameter h corresponds to the bandwdth of the densty estmaton. onvergence s checked by evaluatng the dfference n the mean ponts m(p (t)) - P (t) <h2. We then determne that a user stayed n a tme segment f the resultng mean ponts of the segment are concentrated n the range of a crcle wth a radus of h2. Subsequently, we cluster ndvdual locaton hstores n the spatotemporal doman to obtan stay places and tme ntervals through the followng two procedures. Frst, to obtan stay places, the spatal ponts are clustered by reapplyng the mean-shft procedure to the resultng mean ponts only wthn stay tme segments. Second, we determne a seres of tme segments, n whch the frst one and the last one belong to the same spatal cluster and the ntermedate ones do not belong to dfferent spatal clusters, as the stay tme ntervals. B. Spatal oute Estmaton Based on Stay Places and Trp Patterns Here, we explan how canddate routes are generated from ndvdual stay places detected from locaton hstores n Ds and how the most approprate route s dentfed from them. Because our target of estmatng the trp pattern n ths study s the ordnary moblty route that a user usually takes (.e., he leaves home n the mornng and returns at nght), we ntally determne the locaton of each users home. Here, assumng that users stay home most often, the mode of stay places should smply be home. Subsequently, we connect two consecutve stay places n a trp, and organze tme-ordered trps n a day to a trp pattern. Accordng to ndvdual trp patterns, possble patterns startng and arrvng at the home place are 268

3 exhaustvely enumerated regardless of the occurrence probablty of each trp. At ths tme, we restrct the number of trps between 2 and N trp. In addton, stay places can appear n trp patterns twce or more; for example, a salesman often returns to the offce after vstng some places. Then, to determne the spatal routes of possble trp patterns, we nterpolate each trp wth the shortest path on road and ralway networks by the proposed method n [7]. Here, we only employ the shortest paths and exclude ndvdual dversty of route selecton and traffc nformaton to smplfy the generaton of canddate routes; however, such addtonal nformaton should be ntegrated for further practcal estmaton. As the result, we can prepare canddate routes n the spatal doman. Fgure 2 shows the procedure nvolved n preparng spatal routes. Next, we dentfy the most approprate route for a target day of estmaton through selectng the nearest route n spatal dstance D S between the D trajectory n the target day and canddate routes, both of whch are pont arrays representng trajectores over an entre day (Fgure 3). The spatal dstance D S conssts of the followng parameters: length of the D trajectory N, length of the canddate route N, D poston P, poston n the canddate route P, and ellpsodal dstance dst(p (),P (j)). D S = 2N + 2N N = N = N dst( P ( ), P ( N N dst( P ( N ), P )) ( )) Assumng that both pont arrays have the same length owng to expressng the trajectory over an entre day, the spatal dstance can be calculated as the average dstance between ponts n approxmately the same order. In addton, f consecutve ponts are recorded n Ds, they should be prelmnarly removed for elmnatng bas and for obtanng an accurate spatal dstance. However, the spatal dstance becomes a hgh value n some cases out of our estmaton target, such as routes where the user does not come back home or he/she stays n a place all day. We then defne the threshold value T d for spatal dstance to elmnate exceptons.. Temporal Pattern Estmaton Through eallocatng Duraton of Stay Although the selected canddate route s spatally approprate, t does not contan the occurrence tme of trps and duraton of stays. The duraton of stay s not always the same even f the spatal route and duraton of the trp are dentcal. That s, the duraton of stay should be dversfed to correspond to the occurrence tme of the trps. Therefore, we generate temporal patterns for the selected canddate route by reallocatng the duraton of each stay place wth the unt tme duraton T S. We then obtan the most lkely route as an estmaton result based on the lkelhood functon. By assumng that the D s recorded when a user s wthn radus from a base staton, the lkelhood functon L s (2) Fgure 2. Procedure of generatng feasble routes Fgure 3. Smplfed mage of spatal dstance calculaton defned by the unform dstrbuton wth as follows. The ε value denotes a value suffcently lower than. L N = = log f ( P ( t ), P ( t )) (3) ε dst( P P j f P P j = ( ), ( )) ( ( ), ( )) 2 π (4) ε otherwse Lkelhood can be calculated by comparng canddate routes dversfed temporally wth observed Ds n the target day of estmaton. IV. EXPEIMENTS In ths secton, we explan the expermental results for assessng the proposed methods wth practcal Ds observed n an expermental survey and the consderaton about both results wth hgh and low accuracy. 269

4 A. Summary of Expermental Data From November 28, 20, to December 22, 20, we conducted the expermental survey wth 84 examnees to obtan actvty data: Ds, GPS logs, actvty status data from a web dary, and personal attrbutes va a questonnare. Examnees consented to the prvacy polcy and terms of the experment. In the survey, the average number of accumulated Ds n a day was for each examnee. That s, one telecommuncaton event occurs every 3 mn. Although the Androd applcaton for GPS loggng n every 5 mn generates a telecommuncaton event every 5 mn to send logs, more frequent telecommuncaton events were generated durng ordnary actvtes. For valdatng the proposed estmaton methods mentoned n the prevous secton, we employ Ds and GPS logs n the expermental data. Subsequently, as the baselne data to generate canddate routes for route estmaton, we extract week of D data from November 28 to December 22 and apply the proposed methods to a day arbtrarly selected from the survey perod. For the valdatons, the optonal parameters are set as follows: T = 20 mn, S = 5 mn, h = 4,000 m, h2 =,000 m, N trp = 5 trps, T d = 0,000 m, T S = 30 mn, = 3,000 m, ε = exp(-0). B. Valdaton Method for Estmaton esult Here, we descrbe the method for valdatng the estmated result obtaned wth the methods proposed n the above subsectons. To measure the accuracy by GPS logs obtaned together wth Ds, we defned the followng valdaton functon. The valdaton functon calculates the average dstance E between GPS ponts {P G (t):t = t, t m } and ponts n the estmated result at the same tme slce. Fgure 4. Average number of stay places n a day and total number of stay places durng baselne perod for each examnee Fgure 5. Number of possble trp patterns for each examnee E = m m = dst( P ( t ), P ( t )) G (5). Estmaton esults and onsderaton After applyng the estmaton methods to ndvdual data, we can obtan avalable results about 29 examnees. The route estmaton for the remanng examnees cannot operate correctly because of some exceptons; there are no Ds or GPS logs n the target day of estmaton, or no trp patterns are generated owng to an nsuffcent number of stay places. Therefore, we descrbe the estmaton results and consderaton about 29 examnees as follows. Frst, n the detecton of stay places from Ds, the average number of stay places per day for each examnee s approxmately 2.9, and the average number of total stay places wthn the baselne perod s approxmately 4.4, as shown n Fgure 4. As well as home and offce, where ordnary people usually vst, a few addtonal stay places are detected from actvtes wthn a week. However, because we confgured the bandwdth of densty estmaton to 4,000 m, close range movements such as a short stay at a convenence store nearby the home are concentrated nto a sngle stay place. onsderng the effectve range of the base staton, Fgure 6. Mnmum spatal dstance between canddate routes and D locatons n target day of estmaton Fgure 7. Valdaton results for each examnee 270

5 further detaled detecton s dffcult wth the proposed methods. Fgure 5 shows the number of possble trp patterns generated from stay places for each examnee; there are approxmately 4.9 possble trp patterns per examnee. In comparson wth the total number of stay places, the number of exhaustve trp patterns wthn fve trps s small. We consder that trp patterns to vst certan stay places mght be determned n general. Fgure 6 shows the spatal dstance between the most approprate route nterpolated by the shortest path and locaton hstores of Ds n the target day. The average dstance per each examnee s about 4 km; however, the dstances between examnees are nhomogeneous. We consder that nterpolatng trps wth the shortest path causes ths result because moblty routes are not always the shortest path and depend on personal stuatons such as transportaton cost. Subsequently, the valdaton result shown n Fgure 7 denotes that the average dstance per examnee s approxmately.8 km. The result ndcates that the estmated routes wth the proposed methods are relatvely close to actual trajectores. The ndvdual detals of the accurate result are shown n Fgure 8 and Table. Accordng to Fgure 8, the estmated route n red almost overlaps the GPS logs n yellow except near home, whch s represented by the left-upper stay place n whte. The examnee seems to use the ralway for commutng; however, the estmated route selects a dfferent ralway staton from the actual one on the shortest path nterpolaton. Although the number of detected stay places s only three, as represented n Fgure 8, a more realstc route than the D trajectory can be estmated wth the proposed methods. On the other hand, Fgure 9 and Table 2 show the other naccurate result of the route estmaton. Because the proposed methods employ D hstores for baselne data to generate canddate routes, moblty routes outsde of the D hstores are unavalable for our estmaton methods. In addton, n the spatal dstance calculaton of canddate routes, because both the D trajectory and canddate routes are consdered to represent movement over the entre day, the dfference n the trajectory length causes the estmaton result to worsen. The remanng ssues le n the generaton of a wde varety of canddate routes, and not only n the shortest path, and n the mprovement of the method for determnng the spatal dstance. V. Table. Parameters obtaned n the estmaton procedure for an accurate valdaton result ontent Value Estmated route type alway The number of possble trp patterns 2 The number of Ds n the target day,002 Spatal dstance 6,284 m Log-lkelhood -8,707 The number of GPS ponts n the day 9 Valdaton result (dstance from GPS logs) 982 m Fgure 8. Estmated route, Ds, and GPS logs for examnee of accurate valdaton result Table 2. Parameters obtaned n the estmaton procedure for an naccurate valdaton result ontent Value Estmated route type oad The number of possble trp patterns 3 The number of Ds n the target day 2,370 Spatal dstance 6,889 m Log-lkelhood -46,994 The number of GPS ponts n the day 208 Valdaton result (dstance from GPS logs) 2,789 m ONLUSION AND FUTUE TASKS A. oncluson In ths study, we attempt to estmate a personal moblty route wth spatotemporal consstency over an entre day based on D data. As for the estmaton methods, canddate routes n the spatal doman are exhaustvely generated based on stay places detected from ndvdual locaton hstores n Ds and trp patterns wth the shortest path between them. The spatal route s then determned by dentfyng the nearest canddate route based on shape from the trajectory of Ds durng the target day of estmaton. Addtonally, to consder the dversty of the occurrence tme Fgure 9. Estmated route, Ds, and GPS logs for examnee of naccurate valdaton result 27

6 of the trps on dfferent days, the temporal patterns of the spatal route are generated by reallocatng the duraton of stays. Fnally, the most lkely route n the spatotemporal doman could be estmated by comparng the lkelhood to the observed Ds n the target day. The proposed methods are assessed by Ds and GPS logs obtaned by the expermental survey, wth the result that the average dstance between the estmated routes and GPS logs per examnee s approxmately.8 km. B. Future Tasks Although we employ the shortest path for route nterpolaton of trp patterns, another nterpolaton for generatng a wde varety of canddate routes s necessary. Moreover, n addton to mprovng the method for determnng the spatal dstance between canddate routes, addtonal methods for mprovng the estmaton usng Ds n the target day such as trp segmentaton for smplfyng the spatal dstance evaluaton should be consdered. AKNOWLEDGMENT Ths work was supported by the GENE (Envronmental Informaton) project of the Mnstry of Educaton, ulture, Sports, Scence and Technology (MEXT), Japan, and partally funded by a Grant-n-Ad for Young Scentsts from MEXT. EFEENES [] Y. heng, Mean Shft, Mode Seekng, and lusterng, IEEE Trans. Pattern Anal. Mach. Intell, Vol. 7(8), pp , 995 [2]. Becker,. aceres, K. Hanson, J. M. Loh, S. Urbanek, A. Vasharsvky, and. Volnsky, oute classfcaton usng cellular handoff patterns, Proc. of the 3th AM Internatonal onference on Ubqutous omputng, pp , 20 [3] D. Ashbrook and T. Starner, Usng GPS to learn sgnfcant locatons and predct movement across multple users, Personal and Ubqutous omputng, Vol. 7, , Oct 2003, DOI= [4] J. Froehlch and J. Krumm, oute Predcton from Trp Observatons, Socety of Automotve Engneers (SAE) 2008 World ongress, paper , 2008 [5] F. alabrese, G. D. Lorenzo, L. Lu, and. att, Estmatng Orgn- Destnaton Flows Usng Moble Phone Locaton Data, IEEE Pervasve omputng, Vol. 0, No. 4, pp , 20, DOI= [6]. A. Becker,. aceres, K. Hanson, J. M. Loh, S. Urbanek, A. Varshavsky, and. Volnsky, A Tale of One ty: Usng ellular Network Data for Urban Plannng, IEEE Pervasve omputng, Vol. 0, No. 4, pp. 8-26, 20, DOI= [7] Y. Sekmoto,. Shbasak, H. Kanasug, T. Usu and Y. Shmazak, PFlow: econstructng People Flow ecyclng Large-Scale Socal Survey Data, IEEE Pervasve omputng, Vol. 0, No. 4, pp , 20, DOI= [8] S. Isaacman,. Becker,. aceres, S. G. Kobourov, M. Martonos, J. owland, and A. Varshavsky, Identfyng Important Places n People's Lves from ellular Network Data, Proc. of the 9th Internatonal onference on Pervasve omputng, pp. 33-5, 20. [9] M. A. Bayr, M. Demrbas, and N. Eagle, Moblty profler: A framework for dscoverng moblty profles of cell phone users, Proc. of the Internatonal onference on Pervasve and Moble omputng, Vol. 6, No. 4, pp , 200 [0] T. Horanont and. Shbasak, An Implementaton of Moble Sensng For Large-Scale Urban Montorng, UrbanSense08,