User Category Based Estimation of Location Popularity using the Road GPS Trajectory Databases

Shivendra Tiwari & Saroj Kaushik
Geoinformatica: An International Journal (GIIJ), Volume (4) : Issue (2) : 2014 20
User Category Based Estimation of Location Popularity using
the Road GPS Trajectory Databases
Shivendra Tiwari shivendra@cse.iitd.ac.in
Department of Computer Science and Engineering
Indian Institute of Technology Delhi
New Delhi, 110016, India
Saroj Kaushik saroj@cse.iitd.ac.in
Department of Computer Science and Engineering
Indian Institute of Technology Delhi
New Delhi, 110016, India
Abstract
The mining of the user GPS trajectories and identifying the interesting places have been well
studied based on the visitor’s frequency. However, every user is given the same importance in
the majority of the trajectory mining methods. In reality, the popularity of the place also depends
on the category of the visitor i.e. international vs local visitors etc. We are proposing user
category based location popularity estimation using the trajectories databases. It includes mainly
three steps. First, pre-processing – the error correction and the graph connection establishment
in the road network in order to be able to carry the graph based computations. Second, find the
stay regions where the travelers spent some time off-the-road. The visitors can be easily
categorized for each POI based on the travel distance from the home location. Finally,
normalization and popularity estimation – measure the frequency and stay time of the visitors of
each category in the places in question. The weighted sum of the frequency and stay time for
each category of the visitors is calculated. The final popularity of the places is computed with
values of the pre-configured range. We have implemented and evaluated the proposed method
using a large real road GPS trajectory of 182 users that was collected in a period of over three
years by Microsoft Asia Research group.
Keywords: Trajectory Databases, Trajectory Mining, Popularity Estimation, Region of Interest.
1. INTRODUCTION
The Location Based Services (LBS) are hugely contributing to the next revolution on small
computing handheld such as mobile devices through the mobile network and utilizing the ability to
make use of the geographical position of the device [1]. The location based tour guide is one of
the most useful LBS applications. The tour guides features include route planning, profile based
route optimization, recommending Tourist Point of Interests (TPOI), constraint based tour
planning, location search, map display etc [2]. Mostly the locations are recommended based on
the user ratings. The tour-guide systems require the ratings and popularity measures for the
regions as well in order to recommend the interesting regions known as Region of Interest (ROI).
An ROI can be an individual Point of Interests (POI) at the lowest level or a group of POIs, or an
administrative region such as city, district, villages etc. Such information can help users
understanding surrounding locations, and would enable a better travel guidance experience. The
ROIs are created using different methods as proposed in [1]. It is important to estimate the
popularity of the places in order to disseminate the appropriate location content in such
applications. There are various ways of assessing the popularity of the geospatial locations, such
as user ratings, frequency of the user check-in etc. The average user rating sometimes may not
lead to the estimation of the accurate popularity. It depends on the number of visitors who
participate on rating process, and also deeply depends on their technology awareness, culture,

education etc. The mining of user’s GPS trajectory database is another way of reckoning the
location popularity.
A trajectory is a sequence of sampled locations and time stamps along the route of a moving
object. The analysis of such trajectory data is a critical component in a wide range of research
and decision-making fields. However, it is a challenging problem to analyze and understand
patterns in massive movement data, which can easily have millions of GPS point locations and
trajectory segments. Sometimes the GPS location is far from the actual road and hence it is
important to project them on the road or on the POIs in order to carry about accurate
computations. We have proposed to use map matching techniques to preprocess the GPS
trajectory before actually estimating the popularity of the locations. Map Matching, is the process
of projecting the GPS fixes on the road network graph G = (V, E) [6]. In this paper, we have used
map matching in order to project the inaccurate trajectories on the road network. However, the
trajectories where the GPS points are already map-matched, this step can be skipped. Figure 1
shows the places popularity and their overlap in their popularity index.
FIGURE 1: The place's popularity depending on different scenarios and factors.
The advanced LBS applications requiring the highly accurate information are the main source of
motivation of this work. The ROI based tour guides need to the deliver the appropriately
granulized infotainment data depending on desired level of details. The tourists prefer visiting a
location with several attractions. From the tourist point of view, most of the time the whole region
is considered while planning instead of just one POI. The current popularity estimation systems
mainly evaluate the POIs; and hence, the ratings of the ROIs are unavailable. The popularity of
ROIs is important to be estimated correctly as the granularity of the geographical information, and
dissemination of the content with the desired level of details needs to be assured. There are
multiple challenges of estimating the popularity based on the trajectory database. First,
inaccurate trajectories – the points are sometimes inconsistent, off-the-road and unequal
sampled. The trajectory segments need to be accurate in order to compute the travel distances,
hence they need to be processed in order to reduce the error level. The locations where there is
no digitized map available, therefore, it requires a graph generation process that considers the
average of the multiple trajectories. Second, establishing the relationship between popularity of
the ROI and POIs based on the trajectory visiting trend. The POI visit is not mutually exclusive
with the ROI that encloses a POI. Hence, the POI visit also contributes to the enclosing region’s
popularity. Third, popularity estimation – other than the ratings of the places in order to estimate
the likings, the user visiting frequency also contributes to the trend estimation. We need to
consider the kind of people visiting the underlying while computing their popularity indices.
Our contribution in this paper includes – First, proposed a trajectory analysis method to estimate
the popularity of the geospatial locations. Second, introduce the notion of “who” in order to

categorize the visiting travelers to a location. Finally, we demonstrate and evaluate the method
with the real dataset Geolife provided by Microsoft [8]. Apart from this section as introduction,
section 2 talks about the related work in the trajectory mining area. The data modeling and pre-
processing of the trajectory databases, and determination of the stay points have been discussed
in section 3. Section 4 and 5 establish the algorithms estimating the popularity and experiments
by implementations separately. Finally, the section 6 has concluded the paper.
2. RELATED WORK
Many different methods have been developed for trajectory and movement analysis. In general,
most trajectory analysis methods involve the two steps, first – simplify and generalize each
trajectory; second – compare and group trajectories to find general patterns [5]. Zheng et al in
2009 [9] proposed hierarchical graph based method for mining the interesting locations and travel
sequences from GPS trajectory databases. They model multiple individuals’ location histories
with a tree-based hierarchical graph (TBHG). Based on the TBHG, they defined a HITS
(Hypertext Induced Topic Search) – based inference model, which infers the interest of a location
by taking into account three factors i.e. users travel experience; mutual reinforcement relationship
between travel experience and location interest. Finally, the method mined the classical travel
sequences among locations considering the interests of these locations and users’ travel
experiences. Later in 2010 Zheng et al suggested supervised learning based approach to infer
people’s motion modes from their GPS logs [10]. They also introduced a social networking
service, called GeoLife, which aims to understand trajectories, locations and users, and mine the
correlation between users and locations in terms of user-generated GPS trajectories [11].
Kang et al in 2010 [12] suggested method to mine the spatio-temporal pattern in the trajectory
data – it first finds meaningful regions and extracts frequent patterns based on a prefix-projection
approach from the sequences of these regions. They experimentally proved that the proposed
method improves mining performance and derive more intuitive patterns. Lee et al in 2007 [13,
14] proposed a trajectory clustering method based on the partition and group framework. They
established the importance of discovering the common sub-trajectories in many applications,
especially if we have regions of special interest for analysis. The new framework partitions a
trajectory into a set of line segments, and then, group’s similar line segments together into a
cluster. The primary advantage of the framework is to discover common sub-trajectories from a
trajectory database. Based on the partition-and-group framework, they developed a trajectory
clustering algorithm that consists mainly two phases: partitioning and grouping. The main
advantage of the algorithm is the discovery of common sub-trajectories from a trajectory
database.
Yan in 2009 [15] proposed semantic trajectory analysis based on the statistical computation and
semantic concepts. It involves three major perspectives, i.e. trajectory modelling, trajectory
computing, and trajectory pattern discovery. The early stage of the work has surveyed three types
of modelling requirements for comprehensively explaining trajectories, in terms of geometric
knowledge, geographical knowledge, and application domain knowledge. Zenger in 2009 [16]
proposed the POI recommendation techniques based on the GPS trajectory databases. Zenger
presented a new framework for trajectory-based POI recommendation. The method constructs a
k-truncated generalized suffix tree to represent a historical trajectory database, and use it to
execute exact matching recommendation queries. Two variants are developed, allowing for the
execution of fuzzy matching and order-flexible queries.
3. PROPOSED SOLUTION
Figure 2 shows the high-level component and flow of the trajectory data processing the popularity
estimation including the three major components i.e. preprocessing, determination of the stay
regions, and popularity estimation. The preprocessing module involves improving the trajectory to
make it usable – this step, however can be skipped in case the underlying trajectory has high
quality GPS sequences. The stay region determination step finds out all the regions where the
traveler spent more than a minimum amount of time off-the-road. This step simply produces a list

of rectangular regions that could be POIs as well and hence these regions are useful for
popularity computation in the later stage of the procedure. The popularity estimation step includes
identifying the visiting locations and making a check-in entry into the place registry as discussed
in section 4. In this section we discuss the preprocessing of the trajectory data and the stay
regions determination.
3.1 Data Modeling and Preprocessing
The road GPS trajectory is sometimes inaccurate and has irregular behavior in terms of the time
and the location. Also, the trajectories might have irregular time interval of logging the location.
The other problem is that the trajectories do not consider the connectivity of the roads as a graph.
It is hence difficult to calculate the travel distance only based on the trajectories. In this section,
the preprocessing of the trajectory data is carried out based on the existing map matching and
interpolation techniques [5, 6]. The trajectory is converted into the road network graph which is
aware of the connectivity of the roads and hence it becomes more accurate to compute the
reachability among consecutive points in the trajectory.
Map Matching – The map matching problem is characterized by two objectives – identify the link
traversed by the traveler and find the actual location within that link. Road network map and GPS
data are often enough for post processing map matching. The shortest path algorithm can be
appropriately used for post-processing map-matching. The map matching plays an important role
in putting the user either on the road or in a region based on the users location as discussed in
section 3.2. It is possible that the user stays at the POIs for some time and hence there might not
be the road at all. In such cases, the map matching algorithm has to put the user on the POI so
that it can be concluded that the user indeed visited the POI. So the output of the map matching
is not only the road network graph, but also a list of regions where the user stayed. The stay point
finding method includes the task of map matching in case there is inaccuracy; however, it also
finds the stay regions where the traveler has spent some time. The input to the map marching is
the road network, trajectory database; however the output is the accurate GPS sequence that
guarantees the point being exactly on the road if moving or in any geographical location (if there
is a stay). For the sake of completion, we have discussed the map matching in short, that is
inspired from the work in [6].
FIGURE 2: Overall component diagram for proposed popularity estimation.
Map Generation – In case the digitized geographical map is not available in the area, we need to
generate the map based on the all the trajectory paths. Map generation is a process of creating
the road network based on the trajectory databases. Considering the inherent inaccuracy in GPS
measurements, a circular window is used to smooth/aggregate GPS points and to extract a much
smaller number of representative points for a group of GPS points. A circle with a fixed radius is
placed at each GPS point, whose location will be changed to the average of all the GPS points

covered by the circle. The smoothing process brings the point closer to the road median. If a GPS
point does not have any other point within the specified distance, it will remain in the original
location. We choose a smaller set of new locations as representatives of the original GPS points
to reduce data redundancy and size. If there is no other GPS point within distance criteria to a
GPS point pi, then pi will represent itself. The representative point based map generation method
is inspired by the graph based trajectory analysis method in [5].
Sampling Amplification – GPS points collectively can reveal the road network as we discussed
in the earlier subsection. However, the trajectory also needs to include enough sampling of the
points. If the GPS is retrieved in a long time interval, then it can be heavily inaccurate while
computing the other attributes such as travel speed, distance, stay time etc. Therefore, we need
to insert additional sampling of the points in order to make the trajectory well sampled. The
trajectory interpolation is a method of estimating such points within the trajectory line segments.
The challenge is that this is not a linear interpolation since a straight-line trajectory segment
should be interpolated to follow the curves and turns of the ‘road’. The interpolation achieves
important outcomes i.e. improves the resolution and accuracy; enables accurate location-based
summary statistics; and establishes the topological relations between trajectories.
3.2 Determining Stay Points
The stay points are the regions where the user has spent some time off-the-road. If it is on the
road, it could be due to heavy traffic and hence it needs to be carefully avoided. This process
takes two inputs i.e. road network graph G = (V, E); and user GPS trajectory database T = {T1, T2,
T3… Tm}; where Ti = {<x1, y1, t1>, <x2, y2, t2>, <x3, y3, t3>… <xk, yk, jk>}. It generates a list of
regions R = {R1, R2, R3,… Rk} where user spent some time off the road network for more than a
minimum threshold. The region is a rectangular area with a list of points within it. It will help us
further knowing the POIs user stayed in the later part of the solution.
The extraction of a stay-point depends on two scale parameters, a time threshold (Tthresh) and a
distance threshold (Dthresh). For the points {p5, p6, p7… p17}, a single stay-point s can be regarded
as a virtual location characterized by a group of consecutive GPS points P = {pm, pm+1, … , pn},
where ∀m < i ≤ n, Distance(pm, pi) ≤ Dthresh and |pn.T – pm.T| ≥ Tthresh. Formally, conditioned by P,
Dthresh and Tthresh, a stay point s=(Region, Latitude, Longitude, Tarrival, Tdeparture), where –
P
Latitude.p
s.Latitude
n
mi
i∑=
= (1)
P
Longitude.p
es.Longitud
n
mi
i∑=
= (2)
The region s.Region is the rectangle with center as (s.Latitude, s.Longitude). Equation (1) and (2)
respectively stand for the average latitude and longitude of the collection P. However, s.Tarrival =
pm.T and s.Tdeparture = pn.T represent a user’s arrival and departure times on stay point s included
in rectangular region. These stay points occur where an individual remains stationary exceeding a
time threshold. In most cases, this status happens when people enter a building and lose satellite
signal over a time interval until coming back outdoors. The other situation is when a user wanders
around within a certain geospatial range for a period. In most cases, this situation occurs when
people travel and are attracted by the surrounding environment. As compared to a raw GPS
point, each stay point carries a particular semantic meaning, such as the shopping malls we
accessed and the restaurants we visited, etc.

4. POPULARITY ESTIMATION
The popularity is measured based on the different attributes such as visiting frequency, visitor
category, and stay time. The process includes three steps, i.e. determining the user’s ‘home
location’ and checking-in into the locations in questions. It is interesting to note that our solution is
generic to a single point location i.e. POI as well as a geographic region i.e. ROI. The database in
question may contain both kinds of records which can be evaluated in this section. In order to
keep terminology clear, we use the term ‘location’ for both the POI and ROI.
4.1 User Check-in
The user enters in a location, and spends sometime is known as user check-in. When a user
leaves the location, it is marked as check-out. The check-in location, frequency and the stay-time
creates a pattern that would be helpful determining if it is a ‘home’, ‘routine’ or an ‘interesting’
locations of the other users. A home location for a user is a location that is visited and spent time
there on a regular basis. The visiting pattern such as leaving the place in the morning and coming
back to the same location in the evening in a regular pattern, then it is marked as the home
location. In order to determine the home location, we need to analyze the regions set R for the
individual users. Intuitively, the stay time between two GPS points is the most important attribute
in the trajectory to identify the home location. The POIs under evaluation are stored in the spatial
grid so that the algorithm only search the appropriate grids while looking for the check-in
locations. Only those points are evaluated for the POI check-in only if the user has spent more
than a specified minimum time. The POIs are searched using the adjacent grid based search
(AGBS) method [4]. The stay location where there is no registered place, add the reverse-
geocoded [17] address as the new place. The newly discovered place can be used as potential
POI that is yet to be digitized or it could be a home location for the user that is normally not
digitized as a POI.
Algorithm 1. User-Check-in
Input: Set of POIs to be evaluated denoted by Ŋ; and the stay regions
set R = {R1, R2, R3,… Rk}
Output: Њ =Set of user’s home locations; updated check-in information
for the set of POIs Ŋ;
Steps:
1. Repeat through each region Ri ∈ R
a) Repeat through each point pi in region Ri
i) Search the nearest POI ∈ Ŋ from the point pi using the adjacent
grid search method.
ii) If a valid POI exists within the region Ri, then
- Go to step (iv);
iii) Else
- Create new POI = call reverse-geocode(pi);
- Add the POI place into POI database Ŋ.
iv) Update registry
- Add an entry <user-id, arrival-time, departure-time> in the
POI registry.
- Add the POI into the user’s registry.
2. Evaluate POI database Ŋ:
i) Mark the highest frequency visiting locations and the highest
average stay-time as ‘home’ or ‘routine’ point for this user.
ii) Remove the current user from the POI registry (the user’s home
visit is discounted from the popularity computation)
iii) Remove the home location(s) from the user registry.
3. Update travel distance (from the nearest home location) for the
current user in all the POIs in the POI database registry to
maintain <user-id, arrival-time, departure-time, travel-distance>
4. End.

Algorithm 1 takes the user’s trajectory road network graph and a set of POIs, and the region sets
visited by the user. Once the home location and the visited POIs are identified, the distance from
the nearest home location to the POI is computed and the <user id, stay time, travel distance>
tuple is stored in the POI registry. This algorithm needs to be executed for each visiting user in
the POI database.
4.2 User Categorization, Normalization and Popularity Estimation
Since not all the attributes have the same importance, the analytics hierarchy process can be
used to find the weights for the attributes. The comprehensive decision weights for each
alternative are calculated by the weight sum as suggested in [7]. The global travelers’ visits are
considered more significant in terms of popularity than that of the local visitor. It is important to
note that the user category can be different for a user visiting the different locations depending on
travel distance from the home location.
First, user categorization – the first step is to categorize the visitors in a POI based on their
home location and travel distance. The categorization also uses the home location’s
administrative information in order to decide the border cases. For example, The Golden Gate
Bridge in San Francisco, in USA is closer to the Mexican people living near the Tijuana region
than for the people living in the rest of the states in the USA. It means that only the distance
cannot decide the user category and hence we take the help from the home location’s
administrative information. We propose to divide the users in 5 categories (i.e. global, national,
regional, local, and native) that would have their own weights in the effective weighted frequency
and stay time computation. The distance criterion for each category is application and database
dependent that can simply divide the
Second, weighted sum – for each place the check-in registry is grouped for each user category.
The weighted sum of the frequency and stay time is computed i.e. the frequency Fi and stay-time
Ti of i
th
POI are defined in eq (3, 4). The frequencies (i.e. Fi) and time (i.e. Ti) values multiplied by
the corresponding user category weights and are added up to have an overall weighted sum for
each place. All the frequency and weights are denoted respectively, for global (i.e. Fg, Wg),
national (i.e. Fn, Wn), regional (i.e. Fr, Wr), local (i.e. Fl, Wl) and native (i.e. Ft, Wt). Here ∑W = 1.
Fi = (Fg.Wg) + (Fn.Wn) + (Fr.Wr) + (Fl.Wl) + (Ft.Wt) (3)
Ti = (Tg. Wg) + (Tn.Wn) + (Tr.Wr) + (Tl.Wl) + (Tt.Wt) (4)
Finally, normalization of the results – the stay time and the frequency values are normalized
before actually used in the popularity estimation as shown in eq (5, 6). The overall popularity is
the weighted-sum of the normalized frequency and stay time for the underlying place.
F-Normi = Fi / max
n
i (Fi) (5)
T-Normi = Ti / max
n
i (Ti) (6)
Pi = (F-Normi.Wfreq) + (T-Normi.Wstay-time) (7)
Pi-final = Pi / max
n
i (Pi) (8)
Here, F-Normi and T-Normi are the normalized accumulated frequency and stay time respectively
for the i
th
place. The frequency and the stay time have separate weights as Wfreq and Wstay-time to
compute the weighted popularity Pi in eq (7). Finally, eq (8) computes the normalized popularity
index Pi-final ∈ [0, 1]. The popularity Pi-final can be finally mapped to the context dependent
popularity can be further mapped to the desired range, i.e. ρ = Pi-final * ρmax ʌ ρ ∈ [0, ρmax], ρmax is
the maximum popularity value in the range.

5. IMPLEMENTATION AND EVALUATION
We have used the GPS trajectory dataset that was collected in Microsoft Research Asia’s Geolife
project by 182 users in a period of over three years. This dataset contains 17,621 trajectories with
a total distance of about 1.2 million kilometers and a total duration of 48,000+ hours. These
trajectories were recorded by different GPS loggers and GPS-phones, and have a variety of
sampling rates. This dataset recoded a broad range of users’ outdoor movements, including not
only life routines such as to go home and go to work, but also some entertainments and sports
activities, such as shopping, sightseeing, dining, hiking, and cycling. The majority of the data was
created in Beijing, China [4]. The total data size is approximately 1.55GB which takes around 2
hours 45 minutes to complete parsing the popularity estimation (excluding the preprocessing and
sampling amplification in the trajectory database). With the experiments we have extracted 5617
places as the visited regions. The regions have roughly the rectangular size of 100x80 meters.
Figure 3 shows the data distribution based on the travel distance, collection duration, and the
effective travel duration.
FIGURE 3. a) Trajectory distribution by distance, b) Data collection duration distribution, c) Effective travel
duration distribution.
The demo implementation has been carried out using C++ as the programming language. The
trajectory data have been stored in the flat files; however the spatial grid data structure has been
used to store the places and their user registry. We have used user category weights as 0.39,
0.29, 0.18, 0.09, and 0.05 intuitively in order to compute the weighted sums of the frequency and
the stay times. The global visitors have been assigned higher weights so that their effect is
significantly seen considering that the experimental data has most of the users of the same
country. However, equal weights for frequency and the stay time have been used to compute the
weighted popularity.
FIGURE 4: a) Travel distance to the individual places travelled by the users to visit the location. b) Places
visit frequency distribution per user.

Figure 4 shows the travel distance distribution to the places visited by the travelers during the
data collection period. We have considered all the regions where the user spent more than 10
minutes off-the-road. The first part of the graph shows the travel distance distribution by any user
to visit any place. Mostly the users belong to China who travel within the country under nearly
2000 km away from the home; however, there are some users who travel from US to China back
and forth; hence the travel distance has been significantly higher for some of the instances. In the
second part, the places are plotted with their corresponding visiting frequencies. Although the
number of places visited is good enough; but the frequency in most of the places is low as the
data collection volunteers do not necessarily visit the common places repeatedly. The Figure 5
displays the places visit frequency per user in the first part and stay time distribution by any user
at any place in minutes in the second part of the graph. The high stay time is either due to the
home location or the devices was turned off and restarted after a few days. We have accepted a
maximum of 24 hours of stay at any location.
FIGURE 5: a) The distribution of overall accumulated visit frequency of the places by the individual travelers;
b) The distribution of the travelers stay time across all the places in the <place, user> combination.
S.
No
Place
category
Original
User
Ratings
Total
frequency
Weighted
frequency
Visit
Frequency
based rating
Weighted
frequency
based rating
1 Hotel 5 490 1283.161 4.1727 4.705
2 Hotel 0 307 759.8676 2.5418 2.867
3 Restaurant 0 205 558.1162 2.4835 2.980
4 Shop 0 170 424.6227 2.3625 2.972
5 Restaurant 0 131 355.2668 1.9940 2.292
6
Sights &
Museums
0 122 330.9166 1.8704 2.233
7 Building 5 133 329.8117 2.0202 2.435
8 Restaurant 0 114 297.9203 1.7595 2.656
9 Electronics 0 105 283.7347 1.6374 2.548
10 Snacks 0 126 253.6427 1.9262 2.342
TABLE 1: A comparison of the place rating trend and their estimated popularity.
The most of the places found in the experiment are unrated in the popular mapping and POI data
providers. The major advantage of our approach is that it offers the trend of the place visited by

the overall crowd. Since the user based rating is available on the scale of 0 to 5; we have
mapped the results in the same scale for the uniformity in comparisons in Table 1. It can be easily
noted from the table above that the hotel and the restaurants are the most visited places. Also,
only 2 out of 10 places have been rated by the users. The remaining places did not have any
ratings. The frequency oriented and the weighted frequency based methods have better data
mainly for the unrated places. It is clear that a trajectory based popularity method can be used as
a fallback method for the unrated places or a hybrid approached can be used.
FIGURE 6. a) The distribution of overall normalized frequency based popularity of the individual places; b)
the distribution of normalized weighted popularity.
The travel distance, visit-frequency and stay-time is finally aggregated to the weighted popularity
estimation method in order to compute the final popularity index. Figure 6 shows the normalized
final trend of the popularity indices of the places. We show the comparison based on the
frequency based and weighted method based popularity indices. The first part of the Figure 6
shows the popularity simply based on the frequency; however the second part is the
corresponding weighted popularity ranging within [0, 1]. The travel distances in the experimental
data do not include large variation in the distances as only a few cities of the two countries are
involved in the data collection. However, a clear difference between the popularity distributions
can be seen in a few places where the travel distance was high. Since, the user category does
not vary in the database significantly, the distribution trend is more identical than expected.
6. CONCLUSION AND FUTURE WORK
We have proposed and implemented the popularity estimation method that gives more
importance to the travelers who is taking long way to visit the place. It also considers the fact the
higher visit frequency and stay-time eventually leads to the higher popularity index. The method
is effective enough in the popularity estimation using a rich trajectory database. It is mainly useful
for the places with low user ratings. It can also be used as a fallback method in the
recommendation systems where the initial data is not available. The proposed method has its
own limitations in different aspects. This method might consider the highway side restaurants as
highly popular based on the travel distance of the truck drivers. Our method is mainly focused for
the tourist places; however, it can be easily extended for other kinds of scenarios by simply
considering the context of the use for the locations. The proposed method might reach to
incorrect conclusion in case of the long traffic jams in front of a POI. However, this problem can
be solved by greater GPS accuracy measures. The method does not consider the altitude of the
trajectories in order to estimate the popularity of the POIs on different floors of the same building.
We have estimated the places without considering their POI categories and hence it is good for
the same category places only. However, it is fairly easy to extend the system considering the
specific scenarios and by handling the edge cases in the implementation without loss of
generality. The trajectory data we have used for the experiments have multiple shortcomings
including the small number of users, small number of visited places per user etc.

7. REFERENCES
[1] S. Tiwari, and S. Kaushik. “Modeling On-the-Spot Learning: Storage, Landmarks Weighting
Heuristic and Annotation Algorithm.” In proceedings of CIIA'13 Saida, Algeria, 2013, pp.
357-366.
[2] S. Tiwari, and S. Kaushik. “Fusion of Navigation Technology and E-Learning Systems for
On-the-spot Learning.” In: proceedings of ICWCA’12, Kuala Lumpur, Malaysia. 2012, pp.
72-78.
[3] G. J. Klir and B. Yuan: “From Classical (crisp) Sets to Fuzzy Sets” in Fuzzy Sets and Fuzzy
Logic: Theory and Applications, Prentice Hall PTR (ECS Professional), 1997.
[4] S. Tiwari, and S. Kaushik. “Boundary Points Detection Using Adjacent Grid Block Selection
(AGBS) kNN-Join Method.“ In: proceedings of MLDM’12, Berlin, Germany, 2012, pp. 113-
127.
[5] D. Guoa, S. Liua and H. Jina (2010): “A graph-based approach to vehicle trajectory
analysis.“, Journal of Location Based Services, 4:3-4, 183-199
[6] R. Dalumpines and D. M. Scott (2011): “GIS-based Map-matching: Development and
Demonstration of a Postprocessing Mapmatching Algorithm for Transportation Research.“ In
Lecture Notes in Geoinformation and Cartography, 2011, pp 101-120.
[7] X. Gu and Q. Zhu. “Fuzzy multi-attribute decision-making method based on eigenvector of
fuzzy attribute evaluation space.“ ScienceDirect Decision Support Systems, Volume 41,
Issue 2, January 2006, Pages 400–410.
[8] “GeoLife GPS Trajectories – Microsoft Research“: http://research.microsoft.com/en-
us/downloads/b16d359d-d164-469e-9fd4-daa38f2b2e13/default.aspx
[9] Y. Zheng, L. Zhang, X. Xie and W. Y. Ma. “Mining interesting locations and travel sequences
from GPS trajectories.“ In Proceedings of WWW’09, Madrid Spain, 2009, pp. 791-800.
[10] Y. Zheng, Q. Li, Y. Chen, X. Xie, and W. Y. Ma. “Understanding Mobility Based on GPS
Data.“ In proceedings of ACM UbiComp’08, Seoul, Korea, 2008, pp. 312-321.
[11] Y. Zheng, X. Xie, and W. Y. Ma. “GeoLife: A Collaborative Social Networking Service among
User, location and trajectory.“ In IEEE Data Engineering Bulletin 33(2), 2010, pp. 32-40.
[12] J. Kang, and H. S. Yong. “Mining Spatio-Temporal Patterns in Trajectory Data.“ In Journal of
Information Processing Systems, Vol.6, No.4; December 2010, pp. 521-536.
[13] J. Lee, J. Han, X. Li and H. Gonzalez. “TraClass:Trajectory Classiﬁcation Using Hierarchical
Region based and Trajectory based Clustering.“ In VLDB’08, Auckland, NZ, 2008, pp 1081-
1094.
[14] J. Lee, J. Han, and K.Y. Whang. “Trajectory Clustering: A Partition-and-Group Framework.“,
In proceedings of SIGMOD’07, Beijing, China, June 11–14, 2007, pp. 593-604.
[15] Z. Yan, S. Spaccapietra. “Towards Semantic Trajectory Data Analysis: A Conceptual and
Computational Approach.“, In proceedings of VLDB ’09 Lyon France, August 2009, pp. 24-
28.
[16] Zenger, B. (2009): “Trajectory based Point of Interest Recommendation.“ In Thesis: School
of Computing Science, Simon Fraser University, Canada.

[17] OpenStreetMap (2013): http://nominatim.openstreetmap.org/reverse?lat=37.80948&&lon=-
122.4773&zoom=3 Last accessed on 14-Jul-2014.

User Category Based Estimation of Location Popularity using the Road GPS Trajectory Databases

More Related Content

Viewers also liked (16)

Similar to User Category Based Estimation of Location Popularity using the Road GPS Trajectory Databases (20)

More from Waqas Tariq (20)

Recently uploaded (20)

User Category Based Estimation of Location Popularity using the Road GPS Trajectory Databases