Mining User Mobility Features for Next Place
Prediction in Location-based Services
Anastasios Noulas, Salvatore Scellato, Neal Lathia, Cecilia Mascolo
Computer Laboratory, University of Cambridge
email: firstname.lastname@cl.cam.ac.uk
Abstract—Mobile location-based services are thriving, provid-
ing an unprecedented opportunity to collect fine grained spatio-
temporal data about the places users visit. This multi-dimensional
source of data offers new possibilities to tackle established
research problems on human mobility, but it also opens avenues
for the development of novel mobile applications and services.
In this work we study the problem of predicting the next venue
a mobile user will visit, by exploring the predictive power offered
by different facets of user behavior. We first analyze about 35
million check-ins made by about 1 million Foursquare users in
over 5 million venues across the globe, spanning a period of five
months. We then propose a set of features that aim to capture the
factors that may drive users’ movements. Our features exploit
information on transitions between types of places, mobility
flows between venues, and spatio-temporal characteristics of user
check-in patterns. We further extend our study combining all
individual features in two supervised learning models, based on
linear regression and M5 model trees, resulting in a higher overall
prediction accuracy. We find that the supervised methodology
based on the combination of multiple features offers the highest
levels of prediction accuracy: M5 model trees are able to rank in
the top fifty venues one in two user check-ins, amongst thousands
of candidate items in the prediction list.
I. INTRODUCTION
Understanding human mobility has been a long-standing
subject in academic research due to the multitude of potential
applications. Those range from the better grasp of human
behavior and migration patterns, to the evolution of epidemics
and spread of disease [5], or the understanding of the mech-
anisms that shape social networks [7]. With the introduction
and increasing popularity of location-based services, the op-
portunity to study human movement in a qualitatively novel
setting is provided. Mobile applications such as Foursquare,
where users check in broadcasting their visits to places, allow
us not only to know the geographic coordinates of a user at
a given time, but also the exact places they go. A library,
a cinema or an airport are a few examples amongst the
millions of places which are accessible through these services.
Knowledge about the specific places users visit, which goes
beyond plain geographic coordinates, can be exploited as an
additional dimension to describe human mobility.
In this work we mine user check-in data generated in
Foursquare and study the predictive power that different di-
mensions of the data offer. We formalize the Next Check-in
Problem, where we aim to predict the exact place a user will
visit next given historical data and the current location. The
challenge posed in this context is to rank all the potential target
places in the prediction scenario, which could easily contain
thousands of candidates, so that the actual place visited next
by the user is ranked as high as possible. This represents a
highly imbalanced prediction scenario, where a single correct
instance has to be found (the place a user is going to) amongst
thousands of candidate instances.
Having collected approximately 35 million user check-ins
over a period of 5 months in 2010, taking place over a
set of five million geo-tagged venues, we initially define a
set of prediction features that exploit different information
dimensions about users’ movements: those include information
tailored specifically to an individual user, such as historical
visits or social ties, and features extracted by mining global
knowledge about the system such the popularity of places,
their geographic distance and user transitions between them.
Moreover, we employ a set of features that leverage explicitly
temporal information about users’ movements. We assess the
predictability of individual features and we discover that the
most effective features are those which leverage the popularity
of target venues and user preferences. Next, we combine
the predictive power of individual features in a supervised
learning framework. By training two supervised regressors,
a regularized linear model and M5 model trees, on past
user movements, we demonstrate how a supervised approach
can significantly outperform single features in the prediction
of future user movements, indicating that user behavior in
location-based services is driven by multiple factors who may
act synchronously. Notably, M5 Model Trees rank constantly
one in two user check-ins in the top 50 predicted venues.
II. MOBILE CHECK-IN DATA
10
−2
10
−1
10
0
10
1
10
2
10
3
10
4
10
5
Distance [km]
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
CCDF
(a)
10
0
10
1
10
2
10
3
10
4
10
5
10
6
Time [minutes]
10
−8
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
CCDF
(b)
Fig. 1. Complementary Cumulative Distribution Function of (a) spatial
distance and (b) time elapsed between consecutive check-ins.
Foursquare is one of the most popular location-based ser-
vices, with more than 20 million users as of April 2012.
The Foursquare mobile application allows users to check
in venues via their smartphone: they geolocate themselves,
broadcast their location to their friends, and participate to the
application’s game features. Foursquare users can opt to share
their check-ins publicly to their Twitter profiles. We were thus
able to crawl for publicly-available check-ins via Twitter’s
streaming API. Note that we can only access those check-ins
that users explicitly choose to share on Twitter, although users
have the possibility to set this option as default for all their
check-ins. The resulting dataset contains 35,289,629 times-
tamped check-ins made by 925,030 users across 4,960,496
venues, over a period of 5 months (May 27th to November
2nd 2010). We estimate that this sample holds approximately
20% to 25% of the entire Foursquare user base at the time
of collection
1
. We note that the number of check-ins made
by users is highly heterogeneous: the probability distribution
exhibits a heavy tail, with about 50% of users having fewer
than 10 check-ins. A similar pattern arises when considering
the number of check-ins made in each place: only 10% of
places have more than 10 check-ins.
The probability distribution of spatial distance between
check-ins exhibits a decreasing trend (Figure 1(a)): shorter
distances are more likely to appear. However, longer distances
are still likely, since the decreasing trend seems to obey an
inverse power law (∆r
0
+r)
−β
rather than a faster exponential
decay. A power-law fit [3] to the empirical data suggests
a decaying exponent β = 1.50 and ∆r
0
= 2.87 km. The
distribution of time intervals between consecutive check-ins
is shown in Figure 1(b). Longer intervals are less likely than
shorter ones, denoting that faster sequences of check-ins might
arise, together with long periods of inactivity. This reveals
that users exhibit bursts of check-ins that can be mined to
understand how they choose where to go next. There are two
different trends that become prominent: the first is formed
by consecutive check-ins within 1440 minutes (a day) and
a second, steeper trend when consecutive check-ins happen
across different days. Here, we focus our prediction efforts on
check-ins that happen within 24 hours from the previous one.
III. VENUE PREDICTION IN FOURSQUARE
In this section, we formalize the Next Check-in Problem.
Given the current check-in of a user, we aim to predict the next
place in the city that the user will visit, considering thousands
of candidate venues.
The Next Check-in Problem
We define a set of users U and a set of locations L. Each
check-in c by u ∈ U is defined as a tuple {l, t}, where l ∈ L
represents a venue and t is the check-in’s timestamp. The total
set of check-ins is denoted as C and the set of check-ins for
a specific user u as C
u
. We then formalize the next check-in
prediction problem as follows. Given a user u whose current
check-in is c (to venue l
0
at time t
0
), our aim is to rank the
set of venues L so that the next venue to be visited by the
user will be ranked at the highest possible position in the list.
According to the setting described above, the next check-in
problem is essentially a ranking task, where we compute a
ranking score ˆr for all venues in L.
We constrain the selection of candidate venues to the set of
places L within a given a city. This approach is justified, if
1
http://mashable.com/2010/08/29/foursquare-3-million-users/
one bares in mind that almost 99% of consecutive check-ins
features a distance smaller than 10 kilometers, as shown in
Figure 1(a), suggesting that the vast majority of user activity
in Foursquare occurs within the urban boundary. Further, it
choice allows us to avoid requiring the introduction of distance
as an explicit parameter and we can examine its effect as a
prediction feature in an unbiased way. Finally, the cardinality
of the candidate venue set L in the prediction list varies
from city to city: for the top-33 cities in the dataset we are
experimenting with we note that New York features the highest
number of venues, 43, 681 and Rio de Janeiro trails with 6, 788
observed places.
Mobility Prediction Features.
We now describe in detail the set of prediction features
employed to tackle the next check-in problem. For all cases,
we note as t
0
and l
0
the time and location of the current check-
in respectively. We set t
0
as the current prediction time and we
compute the ranking scores of all features assuming knowledge
up to that time.
User Mobility Features This class refers to features tailored
to the check-ins generated by the user under prediction or by
her social network. Thus, we aim to capture the likelihood that
a user will likely return to a place visited in the past, but also
the likes of the user in terms of types of places.
Historical Visits. By measuring the number of past visits of
user u at a target venue k, we are aiming to assess to what
extent the next check-in of a user is likely to emerge at a place
that has been visited by the user in the past. Formally we have
ˆr
k
(u) = |{(l, t) ∈ C
u
: t < t
0
∧ l = k}| (1)
Categorical Preferences. Another source of information based
on historical behavior is the number of check-ins user u has
performed at a place that belongs to category z. In this way,
we identify the importance of different categories of places
(cinema, coffee shop, football stadium etc.) for a given user
and rank them accordingly:
ˆr
k
(u) = |{(l, t) ∈ C
u
: t < t
0
∧ z
l
= z
k
}| (2)
We note that we subsequently rank venues that belong to
the same category by their popularity in terms of check-in
number. Thus amongst coffee shops for instance, those with
most check-ins are ranked higher.
Social Filtering. Considering a user u and his set of friends
Γ
u
, we rank a target venue k by summing the total number of
check-ins that any friend v of the user has performed at place
k:
ˆr
k
(u) =
X
v∈Γ
u
|{(l, t) ∈ C
v
: t < t
0
∧ l = k}| (3)
Global Mobility Features. Now, we demonstrate how we
can exploit global information about the check-in patterns
of Foursquare users going beyond a specific user and her
social network. In this category we will include popularity
and geographic features together with features that exploit
transitions amongst venues.
Popularity. We define this feature by counting the total number
of check-ins performed by the total set of users U in the dataset
in a venue k:
ˆr
k
(U) =
X
u∈U
|{(l, t) ∈ C
u
: t < t
0
∧ l = k}| (4)
Geographic Distance. To study the effect of geographic dis-
tance in location-based social services we consider the current
location l
0
of user u we measure the distance dist(l
0
, k) to all
other places based on their geographic coordinates. Venues are
subsequently ranked in ascending order.
ˆr
k
(l
0
) = dist(l
0
, k) (5)
Rank Distance. Similarly to geographic distance, we define
rank distance which measures the relative density between
the current place of the user, l
0
, and all other places. Formally,
considering all places l ∈ L we define
ˆr
k
(l
0
) = |{l ∈ L : dist(l
0
, w) < dist(l
0
, k)}| (6)
which in plain words translates to the enumeration of venues
that are geographically closer to l
0
than the destination k.
Our assumption here is that the movement of people is not
based on absolute distance values, but rather by the density of
opportunities or resources nearby.
Activity Transitions. By assuming that the succession of human
activities is not random, as for instance we may visit the
supermarket after work or go to a hotel after landing at
an airport, we are defining the corresponding feature which
enables us to capture this signal in Foursquare check-in data.
Formally, by noting as a tuple, (m, n), the places m ∈ L and
n ∈ L involved in two consecutive check-ins, with z
m
and z
n
being their corresponding categories, we have
ˆr
k
(l
0
) = |{(m, n) ∈ L
c
: z
m
= z
l
0
∧ z
n
= z
k
}| (7)
where L
c
denotes the set of tuples for places involved in
consecutive transitions before current prediction time t
0
.
Place Transitions. By definition of the next check-in problem
we seek to predict consecutive transitions of users across
venues. Thus, we build a feature that directly exploits this
information, by measuring the direct transitions between all
pairs of venues in the city. Accordingly, the rank score of a
target venue k is obtained by enumerating the past transitions
observed by any user from the current location l
0
to location
k, which we formally define as
ˆr
k
(l
0
) = |{(m, n) ∈ L
c
: m = l
0
∧ n = k}| (8)
Temporal Features. Here, we define time aware features that
capture information both on user activity in terms of visiting
categories of places, but also temporal patterns of visits to
specific places. More specifically, given that z
k
denotes the
type of the target place k, we define the Category Hour
Feature APR ACC@10 ACC@50
Random Baseline 0.5 0.0001 0.0005
User Mobility
Historical Visits 0.68 0.30 0.36
Categorical Preference 0.84 0.006 0.05
Social Filtering 0.61 0.17 0.24
Global Mobility
Place Popularity 0.86 0.07 0.16
Geographic Distance 0.78 0.08 0.19
Rank Distance 0.78 0.08 0.19
Activity Transition 0.60 0.03 0.06
Place Transition 0.60 0.17 0.20
Temporal
Category Hour 0.56 0.01 0.02
Category Day 0.57 0.01 0.03
Place Day 0.76 0.07 0.16
Place Hour 0.79 0.09 0.20
TABLE I
MOBILITY FEATURE APR, ACCURACY@10 AND ACCURACY@50.
popularity as the sum of past check-ins at a place of type
z
k
in a given hour h of the day.
ˆr
k
(t
0
) = |{(l, t) ∈ C : z
l
= z
k
∧ tod(t) = tod(t
0
)}| (9)
where tod(t) ∈ [0, 1 . . . 24] returns a value corresponding to
the hour of the day of time t. Similarly, we set Category Day
popularity as the sum of check-ins at a place of type z at a
given hour of a week:
ˆr
k
(t
0
) = |{(l, t) ∈ C : z
l
= z
k
∧ tow(t) = tod(t
0
)}| (10)
where tow(t) ∈ [0, 1 . . . 167] returns a value corresponding to
the hour of the week of time t.
Finally, we also define the temporal check-in activity at
specific venues. We measure the number of check-ins place
k has during a day of the week (Place Day) defined as:
ˆr
k
(t
0
) = |{(l, t) ∈ C : l = k ∧ dow(t) = dow(t
0
)}| (11)
where dow(t) returns the day of the week of time t. A similar
definition follows for the number of check-ins that place k
has at a given hour of a day (Place Hour), aiming to capture
weekly and daily patterns, respectively:
ˆr
k
(t
0
) = |{(l, t) ∈ C : l = k ∧ tod(t) = tod(t
0
)}| (12)
IV. EVALUATING MOBILITY FEATURES
Methodology and Metrics
Given each user check-in eligible for prediction, we have
a set L of candidate places to rank. The features compute
a numeric value ˆr
k
for each candidate venue k, which are
subsequently used to produce a personalized ranking of the
venues. We then denote with rank(k) the rank of venue k,
obtained after sorting in decreasing order all venues in L
according to ˆr
k
. We aim to measure the extent that the future
venue that will be visited is highly ranked by the prediction
algorithms. We use two metrics to measure performance.
First, the Percentile Rank [6] (P R) of the visited place k:
P R =
|L|−rank(k)+1
|L|
. The P R score is equal to 1 when the
place that will be visited next is ranked first and it linearly
decreases to 0 as the correct place is demoted down the list.
The Average Percentile Rank (AP R) is obtained by averaging
across all user check-in predictions: this measure captures
the average normalized position of the correct instance in the
ranked list of instances. We also use prediction accuracy to
assess the performance when using different prediction list
sizes N . In this case, we successfully predict the next check-
in venue if we rank a venue in the top-N places. Average
accuracy is the fraction of successful instances over the total
number of prediction tasks, which we note as Accuracy@N.
Feature based venue prediction
APR Results: The APR results for all features are
presented in Table I. From the class of User Mobility features,
we can distinguish the Categorical Preference feature which
achieves a score 0.84, which is considerably higher than
the Historical Visits (AP R = 0.68) and Social Filtering
(AP R = 0.61). This provides an indication that the types
of places users tend to visit (cinema, nightclub, coffee shops
etc.) can be highly informative about user mobility preferences
and could be employed in mobile applications such as place
recommendation systems. With respect to features mined ex-
ploiting Global Mobility patterns of Foursquare users, Place
Popularity which ranks venues according to the number of
past check-ins is the most promising predictor with an AP R
score that averages 0.86. The Geographic Distance and Rank
Distance attain an average score 0.78, highlighting that spatial
distance is an important factor in the way users decide which
venue to visit next. Continuing in the same class of features,
the Activity Transition and Place Transition features achieve
lower scores with AP R = 0.60, remaining though higher than
the Random Baseline which would achieve 0.50. We close
the AP R score analysis by looking at the performance of
features that exploit Temporal Information about the check-
in patterns of Foursquare users. The Place Hour feature, which
ranks target venues according to the frequency of visits by any
user observed in the past at the current check-in hour, achieves
the highest score, 0.79. The Place Day ranking, which instead
ranks venues by the past number of visits at the day of the
current user check-in, follows closely with an AP R = 0.76,
perhaps due to its lower temporal specificity (day of week
instead of hour of day). Nonetheless, both features signify
that temporal activity around venues constitutes a source of
high quality signal in the venue prediction task. Finally the
Category Hour and Category Day features trail in performance
with scores 0.56 and 0.57 respectively.
The Effect of Prediction List Size: The AP R scores
denote how well, in general, a prediction feature ranks the
next visited venue amongst all candidate venues L. However,
in the context of a real mobile application where a finite set of
places may be recommended to a user, due to interface or other
constraints, one would be interested to examine how prediction
approaches perform when the size of the prediction list N is
limited. We have evaluated all algorithms across various top-N
lengths using the Accuracy@N metric. We show the full set of
results in Figure 2 and we report the results of Accuracy@10
and Accuracy@50 in Table I. The principal observation is that
0 20 40 60 80 100 120
List Size
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
Accuracy
Hist. Visits
Cat. Prefer.
Popularity
Geo Dist.
Rank Dist.
Act. Trans.
Cat. Hour
Cat. Day
Place Trans
Social Filter.
Place Day
Place Hour
Fig. 2. Feature Predictability: Mean Accuracy for all features when they are
being tested on an individual basis for different prediction list sizes N.
features who rank low in AP R can potentially demonstrate
good performance in accuracy terms, in contrast to the results
presented in the previous paragraph,. Overall, the results in
Figure 2 suggest features tailored specifically to User Mo-
bility patterns, such as Historical Visits and Social Filtering
dominate in accuracy for list sizes smaller than N = 60. In
particular, Historical Visits persist over larger list sizes, up
to N = 100. We note that both features had relatively low
AP R scores. On the other hand, features that harvest upon
Global Mobility information, such as Place Popularity or
Geographic Distance fail to achieve high accuracy scores for
small N values. This duality in the performance of the various
predictors can be explained by the fact that some features
can predict exactly the next place a user is going to when,
for instance, the user returns to a previously visited place or
visits places that their friends go to. Nevertheless, the same
features fail to rank appropriately the thousands of previously
unseen venues in the city, thus exhibiting a low AP R score.
We shall see that those heterogeneities in feature performance
will be dissected when we will combine them in a supervised
framework.
Predictability Over Time: We have demonstrated the
overall performance of various features in light of two different
metrics, AP R and Accuracy@N. Another interesting aspect
to consider is how well the different prediction strategies may
perform at different temporal instants. Figure 3 compares the
performance of the various features by showing the temporal
evolution of the AP R score over the week. Overall, the
effectiveness of each feature over time changes: predictions
are more accurate at noon and less accurate in the evening.
This suggests that people might be more habitual during the
day and more likely to alter their patterns and try something
new in the evenings. Interestingly, in the cases of Geographic
Distance and Rank Distance performance is inverted: users
are more likely to cover shorter distances at night between
consecutive check-ins. Further, the variance between the min-
ima and maxima in the temporal results is more prominent for
some features. More specifically, algorithms such as Historical
Visits and Place Transition drop significantly over weekends,
whereas Categorical Preference, Place Popularity and the
distance based features are more stable.
0.5
0.6
0.7
0.8
0.9
1.0
Percentile Rank
Historical Visits
Categorical Preference Place Popularity Physical Distance
Rank Distance
Activity Transition
M T W T F
S S
0.5
0.6
0.7
0.8
0.9
1.0
Percentile Rank
Category Hour
M T W T F
S S
Category Day
M T W T F
S S
Place Transition
M T W T F
S S
Social Filtering
M T W T F
S S
Place Day
M T W T F
S S
Place Hour
Fig. 3. Feature Weekly Predictability: Average Percentile Rank for all features for different hours of a week. Strong daily periodicities are also observed:
notice the yellow circles and red squares which correspond to noon and dinner times respectively.
V. SUPERVISED LEARNING FOR VENUE PREDICTION
In this section, we combine each of the individual predic-
tion features presented previously into a supervised learning
framework. Our aim is to exploit the union of individual
features in order to improve predictions, assuming that user
mobility in Foursquare is driven by multitude of factors acting
synchronously. To predict the next check-in venue of a user we
train supervised models assuming knowledge up to prediction
time t
0
. For every check-in that took place before t
0
, we
build a training example x which encodes the values of
the features of the visited venue (e.g., popularity, distance
from previous venue, temporal activity scores) and whose
label y is positive. Then, we retrieve a negative labeled input
by sampling at random across all other places in the city.
Essentially, we are aiming to teach the model what the crucial
characteristics are that would allow to differentiate places
that attract user check-ins from those which would not. This
method of training a model by providing feedback in the form
of user preference has been established in the past [4] and
corresponds to an effective reduction of the ranking problem to
a binary classification task. Finally, we consider two different
supervised models to learn how feature vectors x correspond
to positive and negative labels: linear ridge regression and M5
decision trees [8].
Results: We are now presenting the prediction results
obtained when we train and test the two supervised learning
models. The M5 trees have the best performance across all
models, with an APR of 0.94 and a clear margin compared to
all single feature prediction strategies that achieve at best 0.86
when venues are ranked according to Place Popularity. On
the other hand, the linear regression model achieves an APR
score equal to 0.81 which ranks it lower than the popularity
and categorical preference features.
If we consider the performance of the models in terms
of prediction accuracy (see Figure 4), we can notice that
M5 model trees dominate with Accuracy@10 equal to 0.31
0.0 0.1 0.2 0.3 0.4 0.5 0.6
Prediction Accuracy
1
10
20
30
40
50
60
70
80
90
100
Recommendation List Size
Lin Regression
M5 Tree
Fig. 4. Average accuracy for the supervised learning algorithms (linear
regression and M5 model trees) for different recommendation list sizes.
and Accuracy@50 equal to 0.51. In the latter case, the next
place visited by the user is on average ranked at the top-
50 positions of the prediction list, which is a remarkable
performance if one considers the multitude of places being
ranked in a city. Compared to the Historical Visits feature
that does best in terms of accuracy, M5 model trees present
constantly better performance: Historical Visits offer good
accuracy scores which however reach an upper bound when
prediction list size N = 10, whereas for larger N values
no improvement is observed. As the reader may notice by
inspecting Figure 4, M5 model trees accuracy performance
ceases to increase rapidly only when N = 100. That means
that their predictive power is not biased by a small set of
candidate venues as in the cases of Historical Visits and Social
Filtering. The linear model presents similar trends in terms of
how its accuracy scores improve relative to list size N but
it fails to achieve high absolute scores, although it still does
better than Historical Visits for N bigger than 50. Overall,
M5 model trees attain peak performance both in AP R and
Accuracy terms, showing not only that that a supervised
approach that combines multiple features is more effective,
but also the fact that this combination is more effective in a
