Understanding predictability and exploration in human mobility
- Andrea Cuttone^{1, 3}Email authorView ORCID ID profile,
- Sune Lehmann^{1, 2} and
- Marta C. González^{3, 4, 5}
Received: 3 August 2017
Accepted: 26 December 2017
Published: 11 January 2018
Abstract
Predictive models for human mobility have important applications in many fields including traffic control, ubiquitous computing, and contextual advertisement. The predictive performance of models in literature varies quite broadly, from over 90% to under 40%. In this work we study which underlying factors - in terms of modeling approaches and spatio-temporal characteristics of the data sources - have resulted in this remarkably broad span of performance reported in the literature. Specifically we investigate which factors influence the accuracy of next-place prediction, using a high-precision location dataset of more than 400 users observed for periods between 3 months and one year. We show that it is much easier to achieve high accuracy when predicting the time-bin location than when predicting the next place. Moreover, we demonstrate how the temporal and spatial resolution of the data have strong influence on the accuracy of prediction. Finally we reveal that the exploration of new locations is an important factor in human mobility, and we measure that on average 20-25% of transitions are to new places, and approx. 70% of locations are visited only once. We discuss how these mechanisms are important factors limiting our ability to predict human mobility.
Keywords
1 Introduction
Billions of personal devices, ranging from in-car GPS to mobile phones and fitness bracelets, connect us to the cloud. These ubiquitous interconnections between the physical and the digital world open up a host of new opportunities for predictive mobility models. Each user of a device produces rich information that allows us to measure their daily mobility routine. This type of knowledge, especially when arising from large numbers of individuals, is expected to impact a wide range of areas such as health monitoring [1], ubiquitous computing [2, 3], disaster response [4], or smart traffic management [5]. Recent contributions to mobility modeling come from computer science [6–8], transportation engineering [9, 10], geographic information sciences [11, 12], and complexity sciences [13–15]. The state-of-the-art for mobility modeling has developed rapidly over the past decade, but further work is needed, especially to tackle the problem of individual predictability.
In the literature, human mobility has been studied using a multitude of proxies (for example call detail records (CDR), GPS, WiFi, travel surveys), and a variety of techniques have been suggested for predictive models, such as Markov chains, Naive Bayes, artificial neural networks, time series analysis. Studies report varying results for the predictive power of these models, with accuracy from over 90% to under 40%.
- 1.
We investigate which factors influence performance in the reported cases of predictability. Showing that the differences in results are driven by the following questions: (a) Does the analysis concern an upper bound of predictability, or actual next-place prediction? (b) How is the prediction problem formulated? E.g. is the goal to predict the next location, or is the goal to identify location in the next time-bin? (c) What is the spatial resolution of the data source? E.g. is the analysis based on GPS vs. CDR data? (d) What is the temporal resolution of the data source e.g. minutes, hours?
- 2.
We quantify the amount of explorations and locations visited only once, and show that these are key limiting factors in the accuracy of predictions for individual mobility.
- 3.
We measure the predictive power of a number of contextual features (e.g. social proximity, time, call/SMS).
- 4.
We study the problem of predictability of human mobility using a novel, longitudinal, high-precision location dataset for more than 400 users.
The rest of the paper is organized as follows. We first provide an overview of related work in the field of predictions of human mobility. Next, we introduce the dataset and describe the preprocessing steps. In the subsequent section we describe the baseline models, and compare their performances. Finally we introduce the exploration prediction problem and report the performance of the predictive models.
2 Related work
In a seminal paper Song et al. [13] investigate the limits of predictability of human mobility, using Call Detail Records (CDR) as proxy for human movement. In their analysis, the authors discretize location into a sequence of places, and estimate an upper limit for the predictive performance using Fano’s inequality on the temporal entropy of visits. Their results show that for a majority of users, this upper bound is surprisingly high (93%). This framework has been further explored to refine the estimate of the upper limit. Specifically, Lin et al. [16] study the effects of spatial and temporal resolution on the predictability limit, Smith et al. [17] consider spatial reachability constraints when selecting the next place to visit, and obtain a tighter upper bound of 81-85%, and Lu et al. [4] analyze the predictability of the population of Haiti after the earthquake in 2010, and find an upper limit of predictability of around 85%. The work described above focuses on estimating an upper limit of predictability for an individual based on an estimate of the entropy their trajectory.
When the topic is actual prediction performance, the most studied models are Markov chains, where the probability of the next location is assumed to depend only on the current location. Markov chains have been applied to a variety of data sets. Lu et al. [18] applied Markov chain models to CDR-based locations in Cote D’Ivore, with a prediction goal of estimating the last location of the day at the prefecture (county) level. Under these conditions the models perform extremely well, reaching an accuracy of over 90%. In [19] the authors apply the Markov models to WiFi traces at Darthmouth campus and find that the best performing model is order 2 and has a median accuracy of about 65-72%. Krumme et al. [20] applied Markov models on shopping locations inferred from credit card transactions, and found that models trained on the aggregate data from a large number of people perform better than individual models. Bapierre et al. [21] applied a variable-order Markov chain to the Reality Mining [6] and Geolife [22] datasets.
Another frequently used type of prediction framework is based on naive Bayes models, where the probability of next location is factorized as independent probabilities for a number of context variables. Gao et al. [23] applied this approach to the Nokia Data Challenge dataset [24] using time and location features, and obtained an accuracy of approximately 50%. Do et al. [25] applied the same technique but used a larger number of features including also SMS, calls and Bluetooth proximity, and obtained an accuracy of approximately 60%. In a subsequent paper [26] the same authors explore a kernel density estimation approach for improving performance.
A number of more complex methods have also been explored in the literature, including non-linear time series [27], Principal Component Analysis [28], Gaussian Mixtures [29] and Dynamic Bayesian Networks [30].
While recent work on predictability has resulted in richer methods and incorporated interesting new features such as social contacts, they have not identified the intrinsic characteristics of human mobility that form the basis for the limitations in predicting the next visited location. In this paper we focus on that aspect, showing that over 53 weeks, individuals visit on average 200 unique locations, of which 70% of them are visited only once. Despite most of the trips taking place among 30% of their (recurrent) locations; we can only predict the occurrence of an exploration with at best 41% of accuracy. Separating visited locations into explorations and recurrent visits may advance the current methods in this area.
3 Materials and methods
3.1 Data description
In this study we analyze a dataset from the Copenhagen Network Study [31]. The project has collected mobile sensing data from smartphones for more than 800 students at the Technical University of Denmark (DTU). The data sources include GPS location, Bluetooth, SMS, phone contacts, WiFi, and Facebook friendships. Data collection was approved by the Danish Data Protection Agency, and written informed consent has been obtained for all study participants.
For this study we focus on the location data, which is collected by the smartphones with a frequency of one sample every 15 minutes. Each location sample contains a timestamp, a latitude and longitude, and an accuracy value. Location is determined by the best available provider, either GPS or WiFi, with a median accuracy of ≈20 meters; more than 90% of the data samples are reported to have an accuracy better than 40 meters. For individual participants, there may be periods missing data. These periods can occur for a number of reasons, for example due to a drained battery, the phone being switched off, the location probe being disabled, or due to software issues. Since we are interested in reconstructing mobility histories without large gaps, we select the longest period that has at least one sample in 90% of the 15-minutes time-bins for each participant. Moreover we consider only participants that have at least 3 months of such contiguous data. We are left with 454 users, with data collection periods of data ranging from three months to one year (Additional file 1 Figure 1).
The data is mainly concentrated in Denmark where the study takes place, but because students use their phones during travel, the dataset spans several other countries as well (Additional file 1 Figure 2).
In this work we are interested in the location prediction task. This task can be broadly stated as follows: given your location history, how well can we predict your future location? The specific details of how this question is implemented have a core impact on the prediction accuracy. Below we investigate how various factors, e.g. spatial and temporal data resolution play a role in determining the reported accuracy for the same underlying dataset.
Because the prediction task can be stated in many different ways, we start the discussion by analyzing different formulations of the problem. In terms of spatial prediction it is possible to discretize space in grid cells, Voronoi cells or define places using a clustering method. In terms of temporal prediction we could decide to predict a location in the next time-bin - within a certain time horizon - or as the next visited place. Below we examine two of the most common problem formulations: next-cell and next-place. In the next-cell formulation we discretize space into grid cells, and we predict the cell in the next time-bin. In the next-place formulation we detect visits to places and we predict the next visited place. The following sections provide details on the two alternative formulations, and show how each formulation affects the prediction task.
3.2 Next-cell prediction
In the first problem formulation, we convert geographical coordinates into discrete symbols by placing a uniform grid on the map and retrieving the grid cell id associated with the coordinates. We define a cell id as a tuple \(( \lfloor \frac{lon}{\Delta lon} \rfloor; \lfloor \frac{lat}{\Delta lat} \rfloor)\), where \(lat\) and \(lon\) are the geographical coordinates, and \(\Delta lat\) and \(\Delta lon\) define the cell-size. Specifically, we start by considering a grid of 50 meters × 50 meters. At each \(\Delta t=15\) minute timestep, we convert the current (lon, lat) into a cell id, producing a sequence of symbols through which we can represent a user’s location history.
We now consider different baseline strategies for next grid cell prediction. For each of the strategies, we perform prediction in an online manner, by training the algorithm on the data up to timestep t, and predicting cell at timestep \(t+\Delta t\). We measure the accuracy as number of correct predictions over the number of total predictions.
3.3 Next-place prediction
We now consider an alternative problem formulation. Instead of predicting the cell in the next timestep, we want to predict only transitions between places, eliminating self-transitions. In order to do so, we convert the raw GPS locations into a sequence of stops at places. A large amount of literature has been dedicated to the problem of place detection, such as methods based on WiFi fingerprints [32], grid clustering [33], and kernel density estimation [34]. In this paper we use a density-based clustering approach such as [35–37].
When detecting places, each user is analyzed individually. We define a ‘stop’ as sequence of locations where the user has been approximately stationary, corresponding to a period of time when the user is visiting a place. In order to extract stops, the location samples \(t_{1},\ldots,t _{n}\) are analyzed sequentially in time, and are grouped into a stop whenever they are within a distance threshold. That is, the distance between position at time t and \(t+\Delta t\) is less than a threshold \(\delta = 50\) meters, roughly corresponding to the GPS accuracy. If a new location sample is farther away than δ, then a new stop is created. This produces a sequence of stops, each one with a centroid calculated as the median of the locations coordinates, and a duration equal to the time between the last location and the first location sample. In order to filter out the short stops that arise, e.g. during commutes, we consider only stops with duration greater than 15 minutes.
We now group stops into places, where a ‘place’ is a group of spatially proximate stops representing a self-contained area such as a building. Example of places are a ‘home location’, a ‘work location’ or a restaurant. In order to group the stops, we apply the DBSCAN [38] clustering to the stops in the geographical coordinate space, using the haversine distance. We set as parameter the grouping distance \(\epsilon = 50\) meters, and min_pts = 2. This distance threshold is set to produce places of the approximate size of a large building (but our results are robust to reasonable variation of this parameter). The result of the DBSCAN clustering is an assignment of a cluster label to each stop, where the label represents the place that the stop belongs to. Finally, in order avoid artifacts due to missing samples or noise, we merge multiple consecutive stops at the same place into one. The process described above converts the raw location history into a sequence of stops at places (Additional file 1 Figures 4 and 5).
The prediction task can now be re-formulated as follows: given a sequence of stops up to step n, can we predict your next place at step \(n+1\)? Notice that a key difference from the cell grid formulation is that in this case there are (by definition) no self-transitions; we are interested in the place changes only.
As before, we start by investigating the upper predictability limit bound. Figure 1 shows how the maximum predictability for the places formulation is peaked at 0.68, significantly lower than what we observe in the next-cell formulation.
Another group [39] has independently analyzed the Copenhagen Networks Study data considering similar questions related to predictability; their conclusions are similar to ours. Our work, however, significantly differs in the sense that we study additional factors that influence predictability, in particular the role of contextual features, and the exploration of new locations.
3.4 Importance of contextual features
We have investigated how the details of the problem formulation impact the reported accuracy for location prediction tasks. We now focus on next-place prediction and study the influence of different contextual features on performance in the prediction task. We formalize the problem as follows: At each step, we aim to compute the most probable next location given the current location, as well as other context variables, such as time of the day or call activity. In other words we wish to compute \(P(\hat{L}\vert c_{1},c_{2},c_{3},\ldots,c_{n})\), where L̂ is the next location, and \(c_{1}\), \(c_{2}\), \(c_{n}\) are the variables representing different contexts.
Description of the features used for the logistic regression models
Feature name | Description |
---|---|
location | location ID |
hour | hour of the day (0-23) |
weekhour | hour of the week (0-167) |
weekday | day of the week (0-6) |
weekend | sat/sun (1) or Mon-Fri (0) |
explore_before | 1 if the previous stop is an exploration, 0 otherwise |
explore_now | 1 if the current stop is an exploration, 0 otherwise |
home | 1 if the current stop is at the home location (most visited place), 0 otherwise |
d_from_home | distance from the current stop to the home location |
sms_received_30min | number of SMS received in the 30 min before the current stop timestamp |
sms_sent_30min | number of SMS sent in the 30 min before the current stop timestamp |
calls_received_30min | number of phone calls received in the 30 min before the current stop timestamp |
calls_sent_30min | number of phone calls sent in the 30 min before the current stop timestamp |
bt_entropy_30min | entropy of Bluetooth devices scanned in the 30 min before the current stop timestamp |
bt_unique_30min | number of unique Bluetooth devices scanned in the 30 min before the current stop timestamp |
As there is no closed form solution for setting the weights, the fit is done by numerical optimization with objective of maximizing classification accuracy.
We model each user separately since our goal is to understand next-place prediction at the individual level. We perform online predictions where we fit the data up to step n, and then predict the next location at step \(n+1\) using a multinomial logistic regression model; classes are simply the set of all possible places visited so far. We fit one model for each individual feature separately, then models combining location and time metadata (hour of the day, hour of the week and day of the week) and then a model with all the features. Models are fit using L2 regularization in order to reduce the number of features. Figure 5(C) displays the accuracy for each of the models, averaged by user.
The Markov chain model baseline is highlighted in red. Using the current location and time features, the logistic regression model outperforms the Markov chain based model. Even using the current location only (which is conceptually very similar to a Markov chain model), the logistic regression shows stronger performance, likely due to the explicit optimization of the model. It is interesting that other context variable such as call and SMS data have little predictive power in this model formulation. The most complex model that considers all features is practically identical in performance to the model using only current location and hour of the week.
The logistic regression model does improve the accuracy over the Markov model, but the absolute value of accuracy still remains low (below 45%). Although the current model could be further fine-tuned, our goal here is not to produce a state-of-the-art prediction model, but instead to investigate the limitations and sources of predictability in human mobility. We also verify the model convergence by studying the accuracy versus the training size, and we find that also in this case the accuracy converges after 50-100 samples (Additional file 1 Figure 6). The logistic regression model shows that (in the current problem formulation) the contextual variables add very little additional predictive power. The next sections will therefore be focused on understanding the factors limiting the prediction accuracy.
3.5 Understanding the set of location states
As we shall see below, another central challenge is not just that our population visits a large number of different places, but also that many new places are discovered over time. We consider a stop at a place as an ‘exploration’ if this place has not been seen in the location history so far for a given user. In other words, this place is being visited for the first time by the user. To express this formally, we consider the sequence of stops \(s_{1}, s_{2}, s_{3}, \ldots, s_{n}\) for each user. We consider a stop \(s_{i}\) as return \((Y=0)\) if \(s_{i}\) has been seen before in the location history, that is there exists a stop \(s_{j} = s_{i}\) for \(1 \leq j < i\). Otherwise we consider stop \(s_{i}\) an exploration \((Y=1)\), that is the place \(s_{i}\) is visited for the first time at step i. For example given a location sequence A B A C B C, the target variable exploration would have values 1 1 0 1 0 0.
We can now estimate the probability of exploration as fraction of explorations over the number of stops. To our surprise, this probability is large: between 0.2 and 0.25 (Figure 6(B)). This implies that most users discover a new place every 4 or 5 stops.
The fact that a large fraction of places have never been seen before poses a challenge for the prediction task, since - by construction - any model that attempts to predict a place from an alphabet of previously visited places will be unable to predict new, unseen symbols. Moreover, another consequence of this exploration is that the pool of possible places constantly grows over time and, given the longitudinal nature of our dataset, ends up being very large. Figure 6(C) and (D) show how the average number of new place explored per week remains approximately constant around 4, and consequently the total number of places keeps growing to hundreds of places. This is a problem for the prediction task, as the number of possible places that the classifier needs to choose from increases constantly.
In fact, if we measure the relation between the number of unique places per user and the performance of the best performing logistic regression model using the Pearson correlation coefficient, we find a negative correlation (\(r = -0.478\), \(p < 0.001\)). Conversely, we find no significant correlation for accuracy with period length or number of stops.
These facts suggest that the exploration phenomenon is a key reason for the relatively low accuracy of mobility prediction tasks at high spatial resolution. Given the importance of exploration, we now introduce a new task in mobility prediction: exploration prediction.
3.6 Exploration prediction
The exploration prediction task can be phrased as follows: given a user’s location history up to step n, will the stop at step \(n+1\) be an exploration or a return?
The first question is: how do we express a useful baseline model for the exploration prediction task? Surprisingly, most literature on human mobility prediction has focused on next location prediction, but has overlooked the exploration prediction problem. Thus, to the best of our knowledge no suitable solution has been proposed for this task. We therefore suggest, as a reasonable baseline, random guessing with probability equal to our prior knowledge of the fraction of explorations: \(P(exploration) \approx 0.2\).
For our main model we use - as before - the logistic regression model with the features used for the next place prediction model. We add two additional features: explore_now and explore_before, which capture if the current stop or the previous stops were explorations, respectively. The intuition for these is that multiple explorations may occur in a row, and therefore the current exploration may increase the likelihood for an exploration at the next stop.
As before, we fit each individual separately, and we perform an online prediction, that fits based on the data up to step n, and predicts exploration at step \(n+1\). We fit one logistic regression model for each of the single features, and a more complex model with all the features at once.
Measuring the performance of these models requires a few considerations. In this case, the classification problem is imbalanced, that is the number of positive cases (exploration) is much smaller than negative cases (return). This implies that accuracy is therefore not a good metric, since a classifier predicting always the most probable class (return) would have good performance, but would not be useful. Instead we use the \(f_{1}\) score, which is the harmonic mean of precision (the fraction of correctly predicted explorations over all predicted explorations) and recall (the fraction of correctly predicted explorations over all true explorations). Figure 6(E) shows the results of the exploration prediction.
As expected, the model with the most complete set of features outperforms the others. Among the single feature models, perhaps not surprisingly, the current location feature has the best performance. This finding can be explained by the role of some places as ‘gateways’ for exploration such as public transport hubs (e.g. a central train station). The individual features that perform also well are the time-related ones, in agreement with the intuition that exploration tends to happen to at specific times of the day or week. The explore_now and explore_before also perform well, suggesting an element of burstiness in the exploratory behavior. If we consider our best performing model, we find that it has average precision of 0.3 and recall of 0.65. Overall the performance of this model is far from perfect, showing that the exploration prediction problem is a challenging one.
4 Discussion
We first showed that when interpreting reported results of predictive performance, there are a number of factors to take into consideration. Central among these is the problem formulation itself. This is simply because predicting the next time-bin is a very different (and much easier) task than predicting the next transition. We showed that the most challenging problem is the next-place prediction, which is arguably the most useful task for practical applications such as travel recommendations. Another issue to be taken into account is the spatial resolution of the prediction, here we show how more coarse spatial precision results in an easier task. Similarly the time resolution also has an effect on the predictive power. We suggest that the factors described in this paper should be taken in consideration as context when comparing results from prediction models.
Other than the factors discussed above, we believe that one further reason for performance differences could be the demographics of the dataset. The data set under study arises from students that have no single workplace but tend to change multiple classes per week, even multiple times per day. Moreover a younger population may have a more irregular schedule and more exploratory behavior. Certainly more work is needed to understand the connection between demographics and predictability. For future research, we suggest considering demographic factors when trying to characterize human mobility, as it has been done, for example, by linking changes in mobility patterns with unemployment status [42].
We also discussed the issue of exploration, and we showed that new places are frequently discovered (typically one in every four or five stops is a new location). Based on this observation, we argued that the mechanism of exploration is an important part of human mobility and plays a role in next-place prediction. Any model that attempts to predict a next place from a set of visited place will fail whenever an exploration occurs. This problem has rarely been addressed in the mobility prediction literature, which almost always assumes that the next place can be determined from the past history. Providing a full solution for next explored place prediction is beyond the scope of this work, and here we simply aim to stress the fact that the prediction of explorations is very different from the predictions to returns to known places. Some previous work on next-place prediction using social information [29, 30] or nearby Points Of Interest [43] may be the starting point for investigating this problem.
In this sense, we raise the question on whether the simple location history is enough for successful next-place prediction. As we have discussed, there are strong regularities both in the sequence of visits, and in the daily and weekly temporal patterns of visitation. However, there are a lot of ‘exceptions to the rules’, where schedules change, plans are canceled, and people run late. We speculate that other channels such as email, social media, calendar, class schedule may be needed for achieving a satisfying accuracy in the prediction task.
Declarations
Availability of data and materials
Data is from the Copenhagen Networks study (http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0095978). Due to privacy consideration regarding subjects in our dataset, including European Union regulations and Danish Data Protection Agency rules, we cannot make our data publicly available. The data contains high spatio-temporal resolution location traces from which is easy to reconstruct personal mobility patterns and infer sensitive personal information.
We understand and appreciate the need for transparency in research and are ready to make the data available to researchers who meet the criteria for access to confidential data, sign a confidentiality agreement, and agree to work under our supervision in Copenhagen. Please direct your queries to Sune Lehmann, the Principal Investigator of the study, at sljo@dtu.dk.
Funding
This work is funded in part by the High Resolution Networks project (The Villum Foundation), as well as Social Fabric (University of Copenhagen). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. There was no additional external funding received for this study.
Authors’ contributions
Designed the study: AC SL MG. Analyzed the data: AC. Wrote the paper: AC SL MG. All authors read and approved the final manuscript.
Ethics approval and consent to participate
Data collection was approved by the Danish Data Protection Agency, and written informed consent has been obtained for all study participants.
Competing interests
The authors declare that they have no competing interests.
Consent for publication
Not applicable.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
- Lane ND, Mohammod M, Lin M, Yang X, Lu H, Ali S, Doryab A, Berke E, Choudhury T, Campbell A (2011) Bewell: a smartphone application to monitor, model and promote wellbeing. In: 5th international ICST conference on pervasive computing technologies for healthcare, pp 23-26 Google Scholar
- Quercia D, Lathia N, Calabrese F, Di Lorenzo G, Crowcroft J (2010) Recommending social events from mobile phone location data. In: Data mining (ICDM), 2010 IEEE 10th international conference on. IEEE Comput. Soc., Los Alamitos, pp 971-976 View ArticleGoogle Scholar
- Aalto L, Göthlin N, Korhonen J, Ojala T (2004) Bluetooth and wap push based location-aware mobile advertising system. In: Proceedings of the 2nd international conference on mobile systems, applications, and services. ACM, New York, pp 49-58 Google Scholar
- Lu X, Bengtsson L, Holme P (2012) Predictability of population displacement after the 2010 Haiti earthquake. Proc Natl Acad Sci 109(29):11576-11581 View ArticleGoogle Scholar
- Çolak S, Lima A, González MC (2016) Understanding congested travel in urban areas. Nat Commun 7 Google Scholar
- Eagle N, Pentland A (2006) Reality mining: sensing complex social systems. Pers Ubiquitous Comput 10(4):255-268 View ArticleGoogle Scholar
- Liao L, Patterson DJ, Fox D, Kautz H (2007) Learning and inferring transportation routines. Artif Intell 171(5):311-331 MathSciNetView ArticleMATHGoogle Scholar
- Zheng Y, Zhou X (2011) Computing with spatial trajectories. Springer, Berlin View ArticleGoogle Scholar
- Arentze T, Timmermans HA (2000) A learning based transportation oriented simulation system. Eirass Eindhoven Google Scholar
- Balmer M, Meister K, Rieser M, Nagel K, Axhausen KW, Axhausen KW, Axhausen KW (2008) Agent-based Simulation of Travel Demand: Structure and Computational Performance of MATSim-t, ETH, Eidgenössische Technische Hochschule Zürich, IVT Institut für Verkehrsplanung und Transportsysteme Google Scholar
- Goodchild MF (2007) Citizens as sensors: the world of volunteered geography. GeoJournal 69(4):211-221 View ArticleGoogle Scholar
- Batty M (2013) The new science of cities. MIT Press, Cambridge Google Scholar
- Song C, Qu Z, Blumm N, Barabási A-L (2010) Limits of predictability in human mobility. Science 327(5968):1018-1021 MathSciNetView ArticleMATHGoogle Scholar
- Simini F, González MC, Maritan A, Barabási A-L (2012) A universal model for mobility and migration patterns. Nature 484(7392):96-100 View ArticleGoogle Scholar
- Schneider CM, Rudloff C, Bauer D, González MC (2013) Daily travel behavior: lessons from a week-long survey for the extraction of human mobility motifs related information. In: Proceedings of the 2nd ACM SIGKDD international workshop on urban computing. ACM, New York, p 3 Google Scholar
- Lin M, Hsu W-J, Lee ZQ (2012) Predictability of individuals’ mobility with high-resolution positioning data. In: Proceedings of the 2012 ACM conference on ubiquitous computing. ACM, New York, pp 381-390 View ArticleGoogle Scholar
- Smith G, Wieser R, Goulding J, Barrack D (2014) A refined limit on the predictability of human mobility. In: Pervasive computing and communications (PerCom), 2014 IEEE international conference on. IEEE, Hungary, pp 88-94 View ArticleGoogle Scholar
- Lu X, Wetter E, Bharti N, Tatem AJ, Bengtsson L (2013) Approaching the limit of predictability in human mobility. Scientific Reports 3 Google Scholar
- Song L, Kotz D, Jain R, He X (2006) Evaluating next-cell predictors with extensive wi-fi mobility data. IEEE Trans Mob Comput 5(12):1633-1649 View ArticleGoogle Scholar
- Krumme C, Llorente A, Cebrian M, Pentland A (2013) The predictability of consumer visitation patterns. Sci Rep 3:1645 View ArticleGoogle Scholar
- Bapierre H, Groh G, Theiner S (2011) A variable order markov model approach for mobility prediction. Pervasive Computing, 8-16 Google Scholar
- Zheng Y, Xie X, Ma W-Y (2010) Geolife: A collaborative social networking service among user, location and trajectory. IEEE Data Eng Bull 33(2):32-39 Google Scholar
- Gao H, Tang J, Liu H (2012) Mobile location prediction in spatio-temporal context. In: Nokia mobile data challenge workshop, vol 41. p 44 Google Scholar
- Laurila JK, Gatica-Perez D, Aad I, Bornet O, Do T-M-T, Dousse O, Eberle J, Miettinen M et al. (2012) The mobile data challenge: big data for mobile computing research. In: Pervasive computing Google Scholar
- Do TMT, Gatica-Perez D (2012) Contextual conditional models for smartphone-based human mobility prediction. In: Proceedings of the 2012 ACM conference on ubiquitous computing. ACM, New York, pp 163-172 View ArticleGoogle Scholar
- Do TMT, Dousse O, Miettinen M, Gatica-Perez D (2015) A probabilistic kernel method for human mobility prediction with smartphones. Pervasive Mob Comput 20:13-28 View ArticleGoogle Scholar
- Scellato S, Musolesi M, Mascolo C, Latora V, Campbell AT (2011) Nextplace: A spatio-temporal prediction framework for pervasive systems. In: Pervasive computing. Springer, Berlin, pp 152-169 View ArticleGoogle Scholar
- Sadilek A, Krumm J (2012) Far out: predicting long-term human mobility. In: Proceeding AAAI’12 proceedings of the twenty-sixth AAAI conference on artificial intelligence Google Scholar
- Cho E, Myers SA, Leskovec J (2011) Friendship and mobility: user movement in location-based social networks. In: Proceedings of the 17th ACM SIGKDD international conference on knowledge discovery and data mining. ACM, New York, pp 1082-1090 Google Scholar
- Sadilek A, Kautz H, Bigham JP (2012) Finding your friends and following them to where you are. In: Proceedings of the fifth ACM international conference on Web search and Data mining. ACM, New York, pp 723-732 View ArticleGoogle Scholar
- Stopczynski A, Sekara V, Sapiezynski P, Cuttone A, Madsen MM, Larsen JE, Lehmann S (2014) Measuring large-scale social networks with high resolution. PLoS ONE 9(4):95978 View ArticleGoogle Scholar
- Kang JH, Welbourne W, Stewart B, Borriello G (2004) Extracting places from traces of locations. In: Proceedings of the 2nd ACM international workshop on wireless mobile applications and services on WLAN hotspots. ACM, New York, pp 110-118 Google Scholar
- Zheng VW, Zheng Y, Xie X, Yang Q (2010) Collaborative location and activity recommendations with gps history data. In: Proceedings of the 19th international conference on World Wide Web. ACM, New York, pp 1029-1038 Google Scholar
- Thierry B, Chaix B, Kestens Y (2013) Detecting activity locations from raw gps data: a novel kernel-based algorithm. Int J Health Geogr 12(1):1 View ArticleGoogle Scholar
- Zhou C, Frankowski D, Ludford P, Shekhar S, Terveen L (2007) Discovering personally meaningful places: an interactive clustering approach. ACM Trans Inf Syst 25(3):12 View ArticleGoogle Scholar
- Zheng Y, Zhang L, Xie X, Ma W-Y (2009) Mining interesting locations and travel sequences from gps trajectories. In: Proceedings of the 18th international conference on World Wide Web. ACM, New York, pp 791-800 View ArticleGoogle Scholar
- Montoliu R, Gatica-Perez D (2010) Discovering human places of interest from multimodal mobile phone data. In: Proceedings of the 9th international conference on mobile and ubiquitous multimedia. ACM, New York, p 12 Google Scholar
- Ester M, Kriegel H-P, Sander J, Xu X (1996) A density-based algorithm for discovering clusters in large spatial databases with noise. AAAI Press, Menlo Park, pp 226-231 Google Scholar
- Ikanovic EL, Mollgaard A (2017) An alternative approach to the limits of predictability in human mobility. EPJ Data Sci 6(1):12 View ArticleGoogle Scholar
- Song C, Koren T, Wang P, Barabási A-L (2010) Modelling the scaling properties of human mobility. Nat Phys 6(10):818-823 View ArticleGoogle Scholar
- Gonzalez MC, Hidalgo CA, Barabasi A-L (2008) Understanding individual human mobility patterns. Nature 453(7196):779-782 View ArticleGoogle Scholar
- Toole JL, Lin Y-R, Muehlegger E, Shoag D, González MC, Lazer D (2015) Tracking employment shocks using mobile phone data. J R Soc Interface 12(107):20150185 View ArticleGoogle Scholar
- Yu C, Liu Y, Yao D, Yang LT, Jin H, Chen H, Ding Q (2015) Modeling user activity patterns for next-place prediction. IEEE Syst J PP(99):1-12 View ArticleGoogle Scholar