Towards an Adaptive Completion of Sparse Call Detail Records for Mobility Analysis

Transcription

1 Towars an Aaptive Completion of Sparse Call Detail Recors for Mobility Analysis Guangshuo Chen, Aline Carneiro Viana, Carlos Sarraute To cite this version: Guangshuo Chen, Aline Carneiro Viana, Carlos Sarraute. Towars an Aaptive Completion of Sparse Call Detail Recors for Mobility Analysis. Workshop on Data Analytics for Mobile Networking, Mar 2017, Kona, Unite States. <hal > HAL I: hal Submitte on 29 Jan 2017 HAL is a multi-isciplinary open access archive for the eposit an issemination of scientific research ocuments, whether they are publishe or not. The ocuments may come from teaching an research institutions in France or abroa, or from public or private research centers. L archive ouverte pluriisciplinaire HAL, est estinée au épôt et à la iffusion e ocuments scientifiques e niveau recherche, publiés ou non, émanant es établissements enseignement et e recherche français ou étrangers, es laboratoires publics ou privés.

2 Towars an Aaptive Completion of Sparse Call Detail Recors for Mobility Analysis (Highly focuse technical solution) Guangshuo Chen, Aline Carneiro Viana INRIA Saclay 1 Rue Honoré Estienne Orves, Palaiseau, France {name}.{surname}@inria.fr Carlos Sarraute Granata Labs Bartolome Cruz 1818 Vicente Lopez Buenos Aires, Argentina charles@granata.com Abstract Call Detail Recors (CDRs) are a primary source of whereabouts in the stuy of multiple mobility-relate aspects. However, the spatiotemporal sparsity of CDRs often limits their utility in terms of the epenability of results. In this paper, riven by real-worl ata across a large population, we propose two approaches for completing CDRs aaptively, to reuce the sparsity an mitigate the problems the latter raises. Owing to high-precision sampling, the comparative evaluation shows that our approaches outperform the legacy solution in the literature in terms of the combination of accuracy an temporal coverage. Also, we reveal those important factors for completing sparse CDR ata, which shes lights on the esign of similar approaches. Keywors Call etail recors, user mobility, human trajectories, location bounaries. I. INTRODUCTION In the past ecaes, the proliferation of personal mobile evices makes Call Detail Recors (CDRs) a very promising source of location information [1]. Collecte by mobile network operators for billing purposes, CDRs ocument the etails about when, where an how mobile phone subscribers generate voice calls or text messages, usually across remarkably large populations. The rich information from CDRs has le to a ramatic increase in mobility-relate stuies, such as ientifying important locations [2], optimizing paging in cellular networks [3], an unerstaning ynamics of human mobility [4]. The sparsity of CDRs often has an averse impact on the epenability of stuy results. Due to the bursty an irregular nature of the communication activities they capture, CDRs are habitually sparse in time, an thus may not recor a user s whereabouts with a stable an consistent frequency. The incomplete mobility information from CDRs causes possible biases on characterizing mobility-relate features [5], [6], [7]. To eal with the sparsity, sometimes heavy filters have to be applie on CDRs to select users having enough mobility information [8]. Data completion aims at filling spatiotemporal gaps in CDRs as much an accurate as possible. It is to locate users continuously in time by leveraging the information of users instantaneous whereabouts. Though it oes not fully conquer the sparsity, as locations logge by CDRs are usually incomplete [7], ata completion can relieve the temporal sparsity of CDRs an problems the latter raises. The legacy solution for completing sparse CDR ata is to hypothesize that a cell tower location ocumente in a CDR is available an representative for a perio (typically one hour) rather than only at a time instant when an activity happens, as use in [8], [9]. This solution is actually a reflection of human nature, i.e., one tens to stay in the vicinity of her voice call places most of the time [10]. A major rawback with the legacy solution is that it always expans all CDRs by the same perio. One this point, previous finings have shown that using a fixe perio at all time is inaequate. In the scenario of etermining whether an when a mobile subscriber stays at home uring the nighttime, Hoteit et al. [6] foun that estimating the home perio aaptively by historical CDRs outperforme the legacy solution with using a fixe perio (10pm, 7am) in terms of accuracy. In the general scenario of completing CDRs uring the aytime, we foun that the legacy solution aforementione might lea to a significant spatial error in our previous work [7], which reporte significantly that the spatial error was positively correlate with the cell size. The stuies above reveal the importance of having an aaptive approach for ata completion for better accuracy. So far, however, there has been little iscussion about this aspect. Although [6] propose an aaptive solution for the scenario of ientifying user s home, it is not a universal esign an oes not consier the environmental information like the cell size. In this paper, we keep on focusing on ata completion for CDRs. We explore (i) what can be extracte from CDRs as features for their completion, an (ii) which features are critical to the esign of a universal aaptive approach. Our results contribute to the effort on reliable CDR ata completion in the following ways. Our investigation is base on two real-worl atasets. Compare with previous GPS atasets, the ataset which we leverage as groun-truth still features high temporal resolution but covers movements of a larger number of users. Details are provie in Sec. II. We propose two aaptive approaches for completing sparse CDR ata an assess their quality on hunres of thousans of CDRs leveraging the groun truth information. They outperform the legacy solution: keep-

3 ing a low spatial error an shortening uncomplete perios. Also, we she light on the main features which are relate to completing CDRs through learning realworl ata. Details are provie in Sec. III. Conclusions are finally iscusse in Sec. IV. II. DATASETS We leverage two atasets collecte from a major cellular operator in Mexico: the target ataset which is compose of CDRs, an the flow ataset to buil groun truth information of user movements. The target ataset contains CDRs of 36, 735 users recore from April 1st to August 31st, On each of these ays, CDRs are collecte uring [10am, 6pm], prevailing working hours. Each CDR provies the etaile information of a user s activity (i.e., a phone call or a text message), consisting of the involve evices (i.e., caller/callee or message sener/receiver, as anonymize ientifiers), the activity time, the routing cell tower location, an the activity uration. The flow ataset is compose of flows collecte uring [10am, 6pm] across the same population as the target ataset 1. Each flow escribes the lifecycle of a TCP or UDP session, an consists of the evice ientifier, the session time, an, particularly, the cell tower location where a session ens. Therefore each user has a fine-graine iscrete trajectory of locations by her flows. Due to the operator s limits of privacy, we can only have three ays of flows (July 19th, 20th an August 9th, 2015). On these ays, we use the flows to construct continuous mobility information as groun truth. For that we complement each iscrete trajectory by expaning each flow from a time instant to a continuous perio, as illustrate in Fig. 2(1). We plot the cumulative istribution function () of the number of recors an the inter-event time in Fig. 1(a) an Fig. 1(b), respectively. The figures reveal that: (i) each user has far more flows than CDRs (95% of users have less than 10 calls but more than 200 flows); (ii) these flows sprea the 8- hour observing perio with a ense temporal coverage (in 95% of cases, a user has two consecutive flows within 100 secons, but only in 20% of cases, has two consecutive CDRs). Overall, the flow ataset contains fine-graine mobility information. Its high temporal granularity ensures flows capture all hanovers of cell towers in the observing perio of each ay, an thus supports the use of trajectories in this flow ataset as groun truth in our analysis. III. CDR DATA COMPLETION In this section, we propose two aaptive approaches for completing CDR ata. The approaches are riven by real ata an aim at filling temporal gaps of unknown locations between consecutive activities. We evaluate their performance 1 The following ata pre-processing steps are carrie out prior to our analysis, in orer to guarantee that every user s movement satisfies an appropriate temporal granularity in the flow ataset. We first apply the recursive lookahea filter on each user s flows to tackle the unesirable effects of celltower oscillation [11]. We then filter out those who have two consecutive flows within higher than 20 minutes. We refer the reaer to [7] for all etails. Flows: Sunay Flows: Monay CDRs: Sunay CDRs: Monay Number of recors (flows or CDRs) per user (a) Number of Recors Flow: Sunay Flow: Monay CDR: Sunay CDR: Monay Inter-event time (secons) (b) Inter-event Time Fig. 1. (a) of the number of CDRs (as ashe lines) an flows (as soli lines) per user. (b) of the inter-event time between two consecutive CDRs (as ashe lines) an flows (as soli lines). Estimate CDR Location Bounary Actual Time: Flows Cell: t1 -t A B C D t2 t3 t4 t5 A A B C D t = (t2+t3) Fig. 2. A emo of (1) constructing the groun truth an (2) estimating a location bounary by the static approach: (1) Suppose five consecutive flows recore at time t 1,..., t 5 an at cell locations A, B, C, D. The two flows at t 1 an t 2 are merge as they are observe consecutively at the same cell A. Each transition between two cells is assume to occur at the mi-time of corresponing consecutive flows. In this way, a continuous trajectory is built. (2) The static approach estimates a fixe-perio location bounary ( t, +t ) attache with a CDR at time t CDR at the cell C, so as to assume the user remains at the cell C uring this perio, while actually she moves from the cell B to D, inicating that only a sub-perio is accurate in this location bounary. by comparing with the legacy solution introuce in Sec. I which we refer as the static approach in the following. The static approach is to hypothesize that a user always stays in the corresponing cell uring a perio centere at the time of each CDR, as illustrate in Fig. 2(2). By the static approach, we present the iea of location bounary. Each location bounary contains a perio corresponing to the completion of a CDR representing that a user s whereabout uring this perio. In the static approach, all CDRs are complete by symmetric location bounaries of the same size ( t, +t ). The perio in a location bounary shoul be estimate accoring to the very situation of the activity. For instance, a location bounary eserves a large t if the user is walking when making a call, but fits a small t if she is on a highspee train. The arbitrary etermination of t in the static approach leas to a significant spatial error in practice ue to the complexity of a user s realistic behavior [7]. With regars to this, we introuce two novel aaptive approaches, where t (or t (s), t(e) ) in a location bounary is aaptively etermine by its own CDR, unlike the static approach which uses a unifie threshol. They are (i) the sym-aaptive approach makes a CDR into a symmetric location bounary as ( t, +t ) an (ii) the asym-aaptive tcdr +t

4 approach makes a CDR into an asymmetric location bounary as ( t (s), +t(e) ). In the following, we introuce how a location bounary is estimate in the two aaptive approaches in Sec. III-A. After that, we compare all the three approaches from two perspectives: accuracy an coverage, iscusse in Sec. III-B. A. Determining aaptive location bounaries Observable factors from CDRs: Leveraging a large number of a user s CDRs collecte uring the long-term observing perio, we can learn her behavior for ientifying a location bounary though the following factors in three categories: 1) Event-relate factors: These are the metaata of a CDR, incluing the activity s time, type (call/message) an uration 2. 2) Long-term behavior factors: The raius of gyration (URg) of a user, the number of a user s locations (ULoc) appearing in the observing perio, an the number of a user s active ays (UDAY). These factors characterize a user by giving senses of (i) her long-term mobility an (ii) her habit on generating calls an text messages, compute by her CDRs prouce uring the 5-month observing perio. 3) Location-relate factors: The first factor in this category is relate to the cell size 3, i.e., the average call raius (CR). Since we have no knowlege of the actual cell coverage, we assume a homogeneous propagation environment an an isotropic raiation of power in all irections at each cell tower, so that we are able to roughly estimate each cell s CR using a composition of Voronoi cells extracte from CDRs which covers the area, as in [7]. The rest of the factors escribe the location where the activity happens regaring its importance to the user. For that, we learn from the algorithm presente by Isaacman et al. [2], which is esigne to etermine prominent locations which the user usually spens a large amount of time an/or visits frequently. Their algorithm firstly clusters all locations which appear in a user s trajectory of CDRs, an then ientify for each cluster whether a cluster is important by measuring these observable factors erive from each cluster in the following, as use in our work: (i) the number of ays on which any cell tower in the cluster was contacte (CDay); (ii) the number of ays which elapse between the first an the last contact with any location in the cluster (CDur); (iii) the sum of the number of ays cell towers in the cluster were contacte (CTDay); (iv) the number of cell towers the cluster (CTower); (v) the istance from the activity location to the centroi of the cluster (CDist). Training an estimating: We evelop our approaches by firstly training our moel with a set of 65, 791 CDRs collecte on August 9th. The other two ays, i.e., July 19th an 20th, on which we have the groun-truth information are then utilize into the testing phase. In this way, the moel is traine having as input the above escribe factors an the location bounaries extracte from the groun-truth information. More 2 For this attribute, the uration of a text message is set to 0 secon. 3 We exclue the antenna-relate information such as the transmit power or RF propagation environment, which a mobile operator rarely provies. URg CR Time UDay ULoc Dur CTDay CDist CDur CDay CTower Type Relative Importance Fig. 3. Relative Importance of features in etermining accurate location bounaries. specially, given a CDR, its user an location, we erive a vector (x 1,, x n ) from the above factors, such as: (i) the categorial factor type is converte to two binary features by one-hot encoing; (ii) the time is converte to two separate time ifferences (in secons) from the activity time to 10am an to 6pm; (iii) other factors are use as values they are. In the sym-aaptive approach, this vector an the symmetric location bounary ( t, +t ) of each activity extracte from the groun-truth information are provie as input to a moel traine by Graient Boosting Regression Trees (GBRT) [12], uner the square loss function. Once the training phase is complete, the testing phase consists in estimating the interval ( t, +t ) by feeing the vector erive from 136, 562 CDRs collecte on the other ays (July 19th an 20th) into the regression formula. In the asym-aaptive approach, the factors vector an the two limits of an asymmetric location bounary, i.e., t (s) an t (), of each activity extracte from the groun-truth information are separately provie as input resulting in two traine GBRT moels. The training phase is the same as in the sym-aaptive approach but works inepenently on the two moels. Fig. 3 shows the relative importance of factors with respect to the estimation of a location bounary after the training phase of the aaptive approaches. This figure allows us rawing the following main conclusions, vali for both approaches. We notice the three most important factors: the activity s time, the cell raius, an the raius of gyration. This inicates that for a cell, how long a user stays insie mainly epens on its size, the precise time the activity occure, an the user s long-term mobility. Surprisingly, the activity s type is the most pointless factor, inicating that knowing whether a user generates a call or a message is useless in etermining a location bounary. B. Coverage an accuracy To valiate the location bounaries given by the three previous iscusse completion approaches, we again use the CDRs collecte on July 19th an 20th, on which we have the groun-truth information. Note that our approaches are applicable to CDRs on other ays, though there is no groun-truth information. Hereafter, we use the ays having groun-truth information only for comparison an valiation reasons. For

5 the static approach, t is set to 5/15/30/45/60/90/120 minutes for location bounaries, respectively. Aitionally, since the goal of ata completion is to fill the temporal gaps, we compare the three approaches in terms of accuracy an coverage. For that, we use the following measures to quantify the complete CDRs: Spatial error: The average cumulative error in the istance over time uring a location bounary, quantifies accuracy in terms of activity. Coverage: The filling rate, efine as the ratio of the covere uration (i.e., the sum of the perios of a user s location bounaries) to the observing perio, represents to what egree a user s CDRs are complete. Accuracy: The ratio of the accurate uration (i.e., the sum of all accurate sub-perios) to the covere uration (i.e., the sum of all perios) on a user s location bounaries, quantifies accuracy in terms of the user. Intuitively, an ieal ata completion approach shoul cover the observing perio as much an precise as possible, i.e., satisfying high accuracy an coverage simultaneously. Fig. 4(a)(b) plots the istribution of the spatial error over all location bounaries. It confirms that the spatial error increases as t becomes larger when using the static approach, which is reveale in our earlier work [7]. More importantly, the performance of the two aaptive approaches is nearly as goo as the static approach with t = 30 minutes in terms of the spatial error. To further compare the approaches in terms of the combination of accuracy an coverage, we plot in Fig. 4(c)() the mean of accuracy versus mean of coverage over the observing users. We notice that using a large t contributes to enhancing coverage with reucing accuracy as a price, inicating that the static approach cannot achieve high accuracy an high coverage together. However, the two aaptive approaches show spleni performance. The sym-aaptive approach reaches the same level of accuracy as the static with t = 30 minutes. Yet it has a better temporal coverage (approximately as goo as the static with t = 90 minutes). This inicates that the symaaptive approach can complete more time while it can still ensure accuracy. As to the asym-aaptive approach, it performs better in terms of coverage with losing a small egree of accuracy, compare with the sym-aaptive approach, an still outperforms the static. Overall, we see a clear avantage of the sym-aaptive approach over the others in Fig. 4, as it achieves the best combination of fair accuracy an high temporal coverage. IV. CONCLUSION In this paper, we focuse on ata completion an propose two novel ata-riven approaches which utilize multiple factors of a user s behavior an etermine location bounaries aaptively for completing CDRs. The comparative evaluation showe that the propose approaches outperforme the legacy mean of coverage 0.9 t = 5 min t = 30 min t = 60 min t = 120 min sym-aaptive asym-aaptive Spatial Error of Location Bouary in ( t, +t) (km) (a) Spatial Error: Sunay ±120min ±90min Static Sym-Aaptive Asym-Aaptive ±60min ±45min 0.3 ±30min ±15min 0.1 ±5min mean of accuracy (c) Accuracy/Coverage: Sunay mean of coverage 0.9 t = 5 min t = 30 min t = 60 min t = 120 min sym-aaptive asym-aaptive Spatial Error of Location Bouary in ( t, +t) (km) 0.3 (b) Spatial Error: Monay ±120min ±90min ±60min ±45min ±30min Static Sym-Aaptive Asym-Aaptive ±15min 0.1 ±5min mean of accuracy () Accuracy/Coverage: Monay Fig. 4. of the spatial error of location bounaries compute on (a) Sunay an (b) Monay; mean of accuracy versus mean of coverage per user on (c) Sunay an () Monay, across the static, sym-aaptive an asym-aaptive approaches. solution in the literature. Further research shoul be one to investigate a heuristic approach which oes not rely on the contextual information. ACKNOWLEDGMENT The authors woul like to thank GranData for proviing the ata use for the experiments. This work was supporte by the EU FP7 ERANET program uner grant CHIST-ERA MACACO. REFERENCES [1] D. Naboulsi, M. Fiore, S. Ribot, an R. Stanica, Large-scale mobile traffic analysis: a survey, IEEE Communications Surveys & Tutorials, vol. 18, no. 1, pp , [2] S. Isaacman, R. Becker, R. Cáceres, S. Kobourov, M. Martonosi, J. Rowlan, an A. Varshavsky, Ientifying important places in peopleś lives from cellular network ata, in Pervasive computing, 2011, pp [3] H. Zang an J. C. Bolot, Mining call an mobility ata to improve paging efficiency in cellular networks, in ACM MobiCom 2007, 2007, pp [4] M. C. González, C. A. Hialgo, an A.-L. Barabási, Unerstaning iniviual human mobility patterns, Nature, vol. 453, no. 7196, pp , Jun [5] G. Ranjan, H. Zang, Z.-L. Zhang, an J. Bolot, Are call etail recors biase for sampling human mobility? ACM SIGMOBILE Mobile Computing an Communications Review, vol. 16, no. 3, pp , [6] S. Hoteit, G. Chen, A. Viana, an M. Fiore, Filling the gaps: On the completion of sparse call etail recors for mobility analysis, in ACM CHANTS 2016, 2016, pp [7] G. Chen, S. Hoteit, A. C. Viana, M. Fiore, an C. Sarraute, Relevance of Context for the Temporal Completion of Call Detail Recor, INRIA Saclay, Research Report RT-482, Nov [Online]. Available: [8] C. Song, Z. Qu, N. Blumm, an A.-L. Barabási, Limits of Preictability in Human Mobility, Science, vol. 327, no. 5968, pp , Feb

6 [9] H. H. Jo, M. Karsai, J. Karikoski, an K. Kaski, Spatiotemporal correlations of hanset-base service usages, EPJ Data Science, vol. 1, pp. 1 18, [10] M. Ficek an L. Kencl, Inter-Call Mobility moel: A spatio-temporal refinement of Call Data Recors using a Gaussian mixture moel. IEEE INFOCOM 2012, pp , [11] M. A. Bayir, M. Demirbas, an N. Eagle, Mobility profiler: A framework for iscovering mobility profiles of cell phone users, Pervasive an Mobile Computing, vol. 6, no. 4, pp , [12] J. H. Frieman, Greey function approximation: a graient boosting machine, Annals of statistics, pp , 2001.