Enabling Semantic User Context to Enhance Twitter Location
Prediction
Ahmed Galal and Abeer El-Korany
Department of Computing Science, Faculty of Computers and Information, Cairo University,
Cairo, Egypt
Keywords: Dynamic User Modeling, Location Prediction, Semantic Context, Topical Interest, User Behavior.
Abstract: Prediction of user interest and behavior is currently an important research area in social network analysis.
Most of the current prediction frameworks rely on analyzing user’s published contents and user’s
relationships. Recently the dynamic nature of user’s modelling has been introduced in the prediction
frameworks. This dynamic nature would be represented by time tagged attributes such as posts or location
check-ins. In this paper, we study the relationships between geo-location information published by users at
different times. This geo-location information was used to model user’s interest and behavior in order to
enhance prediction of user locations. Furthermore, semantic features such as topics of interest and location
category were extracted from this information in order to overcome sparsity of data. Several experiments on
real twitter dataset showed that the proposed context-based prediction model which applies machine learning
techniques outperformed traditional probabilistic location prediction model that only rely on words extracted
from tweets associated with specific locations.
1 INTRODUCTION
Online social networks are very popular platforms
that allow users to publish different types of contents
that express their interest and ideas. Earlier in social
networks, most of these published contents were
merely textual posts, photos or videos. While today,
online social networks have introduced location-
based services that allow user to publish geo-location
information from different locations and at different
times of the day. Typical location-based social
networking sites allow users to “check in” at a
physical place, share the location with their online
friends, rate, and provide tips on the visited locations.
Recently, most well-known online social networks
like Facebook and Twitter have incorporated the
location services to allow their users to tag posts to
locations. Others location-based social networking
services available are: Foursquare, Gowalla, Google
Places and Yelp.
The heterogeneous data in location-based social
networks contain spatial-temporal social context and
present new challenges and opportunities for further
analysis. Those information are associated with users
and was used to dynamically model users in social
network (Galal and ElKorany, 2015) since both
content-based and location-based information are
changing over time. Generally, users profile attributes
could be classified into dynamic attributes and static
attributes. Dynamic time-tagged attributes such as
topics used in posts and comments as well as location
check-ins which are used to represent user’s interest
and behavior respectively. While, static attributes
represent information that rarely change with time
such as demographic attributes. Recently, most of
research in social network analysis utilize those
dynamic attributes to further be used in multiple
social network analysis tasks such as recommender
systems (Bobadilla et al., 2013; Abel et al., 2011),
expert identification (Kleanthous and Dimitrova,
2008), location prediction (Chandra et al., 2011;
Cheng et al., 2010; Ye et al., 2013), link prediction
(Quercia et al., 2012) and similarity measurement
between users (Galal and ElKorany, 2015; Lee and
Chung, 2011).
Moreover, semantic information extracted from
those dynamic attributes like topics or the category of
location are used in automatic mapping of users to
their “key” visited locations of interest (e.g., home,
work, leisure). This mapping is done based on their
online social presence and has been of great interest
for the research community (Mahmud et al., 2014;
Galal, A. and El-Korany, A.
Enabling Semantic User Context to Enhance Twitter Location Prediction.
DOI: 10.5220/0005749502230230
In Proceedings of the 8th International Conference on Agents and Artificial Intelligence (ICAART 2016) - Volume 1, pages 223-230
ISBN: 978-989-758-172-4
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
223
Ryoo and Moon, 2014). Furthermore, relationships
between multiple dynamic attributes have emerged
e.g. textual posts tagged with location check-in,
images tagged with location check-in or images that
contain textual description or comments. These
relations between different dynamic attributes are not
always realized and ignored. Although they could be
mined and analyzed to add a big advantage by either
to increase the accuracy (Galal and ElKorany, 2015)
or better understand or represent the dynamic
behavior and interest of users (Li and Chen, 2009).
Location prediction task attracted a significant
amount of researchers (Wang and Prabhala, 2012).
Being able to predict people’s future location and
hence, marketers could decide to do additional
advertising at certain events or during particular TV
shows. In this paper, a novel framework for
predicting the category of user’s current location
using her/his geo-tagged Twitter activity and
behavior such as topic of interest and category of
previously visited places is proposed. Historical
knowledge representing user activities while visiting
key locations, as well as time of posting contents is
used to enhance location prediction. This historical
knowledge is represented by topics extracted from
tweets that are posted while visiting similar location
associated with the posting time of those tweets. By
linking twitter user account with foursquare accounts,
relationship between time, topics and location of
users have been exploited.
The rest of the paper is organized as follows. In
Section 2, we discuss the related work in prediction
of user’s behavior and interest. In Section 3, we
explain the main components of the proposed location
prediction framework. Results and accuracy
evaluation on twitter dataset is discussed in Section 4.
Finally, we draw our conclusion and discussed
intended future work in Section 5.
2 RELATED WORK
Prediction of user's behavior and interests is an active
area of research in social network analysis. Most of
the proposed frameworks in these researches rely on
user’s published contents to predict his behavior or
interest. A prediction framework proposed in (Jamali
and Rangwala, 2009) that relies on machine learning
to predict the popularity of posts in Digg based on
comments statistics, users’ interest which is
represented by rate of commenting, users’ feedback
on a post, and by utilizing the users’ community
structure. Another framework proposed by
(Weerkamp and De Rijke, 2012) that predicts future
users activities based on terms extracted from tweets.
This framework predicted tonight activities only
without considering other timeframes within the day
or whether today is a weekend or a weekday is.
Nowadays, one of the most popular features to be
predicted is user’s locations. These locations can be
predicted using several ways either by analyzing
dynamic attributes such as posts or tweets or by
studying the history of users’ movements.
There are two major goals in location prediction.
The first one is to predict user’s actual physical
location such as the current city (Chandra et al., 2011;
Cheng et al., 2010). While the second aims to predict
the semantics of the location such as location
category or type (Ye et al., 2013).
(Chandra et al., 2011) Predicted city level
location of users based on the probability distribution
of terms with respect to locations. These terms are
extracted from tweets and reply-tweets. Another
probabilistic based framework that predicts city level
location of users is proposed in (Cheng et al., 2010).
The main contribution in this framework is that they
have managed to handle the sparsity of tweets and the
nonstandard vocabulary that exist within the tweets.
The previously mentioned city level prediction
frameworks are different than our proposed
framework in that they did not consider the effect of
time on user published content as well as the semantic
relation between tweets and locations from where
users tend to post specific content. On the other hand,
the proposed framework predicts the category of
location which is more significant when considering
users actual location as users who are living in
different countries or cities and may not be able to
visit the same physical location.
Modeling of human mobility and a probabilistic
framework for location prediction is proposed in (Cho
et al., 2011). This framework relies on past user
check-ins extracted from location based social
networks and cell phone location tracing. (Ye et al.,
2013) Proposed a framework which predicts the
category of the user activities and the most likely
location to be visited. By using mixed hidden Markov
model to generate the activities category distribution
using user’s history of movement, activities and
which is used further in location prediction.
Unlike all of the above mentioned works, our
proposed research work makes use of some of these
findings while also going further. The proposed
framework predicts the category of user’s current
location (not the actual physical location) using topics
extracted from tweets that have been posted during
his stay in the location. Furthermore, taking into
considering the assumption that people tends to visit
ICAART 2016 - 8th International Conference on Agents and Artificial Intelligence
224
some places on specific times or specific days,
posting time of their tweets is considered also as a key
factor in prediction of the category of user’s location.
3 PROPOSED CONTEXT-BASED
LOCATION PREDICTION
FRAMEWORK
Location prediction is one of the most hot research
topics in social network analysis. Current research in
location based social network mainly focuses on two
tasks: 1) predicting a user’s home location; and 2)
predicting a user’s location at any time. The former
task considers predicting the static home location of a
user, while the latter considers more about predicting
a user’s moving trajectories (Gao & Liu, 2014). The
proposed framework lends itself to second category
of prediction tasks which aim to predict the category
of user’s current location.
Some researchers have considered the correlation
between specific terms in tweets and their
corresponding locations (Cheng et al., 2010; Hecht et
al., 2011). Thus, for the purposes of enhancing
prediction of user’s current location, we propose a
context-based model that integrates both content-
based and location-based attributes to investigate the
relationship between the published posts and the
user’s check-in behavior and their variation over
time. The proposed framework can be divided into
two major components; the first component is
responsible for identifying and modeling of users’
context such as locations and topics, while the second
component is the prediction engine which will be
explained in details in the following subsections.
3.1 Modeling of User Context
The first component in the proposed framework is
responsible for creating the dynamic user model
which captures user’s topical interests based on
his/her geographical locations. Users from different
countries or cities can visit similar places that belong
to similar categories (Galal and ElKorany, 2015).
Thus, our proposed context-based prediction model
uses novel temporal features not used by any existing
work. According to (Dalvi et al., 2012), who studied
the problem of matching a tweet to an object, where
the object is from a list of objects in a given domain
(e.g., restaurants). Their model is based on the
assumption that the probability of a user tweeting
1
https://developer.foursquare.com/categorytree
about an object depends on the distance between the
user’s location and the object’s location. Such
matching can also geo-locate tweets and infer the
present location of a user based on the tweets about
geo-located objects. Accordingly, we assumed that a
combination of geographical information and topic
model could be used to discover user’s current
location. Furthermore, based on the hypothesis that
users tend to post content related to the location they
currently visit, we assumed that half an hour time
frame to stay in specific place increases the likelihood
that user start posting text related to the current place
(different time slots values were used till the model
become stable). For example, while a user is sitting in
a restaurant waiting for menu, she/he usually post
tweets related to type of food she/her prefer or post
about trips while waiting for her/his plan in an airport.
Therefore, we collect location-based users’ tweets
within half an hour after detecting a location check-in
done by the user using foursquare social network. The
topical interests will be represented as a set of vectors
for each user such that one vector is used to represent
set of topics posted by a user in specific location’s
category visited by him/her. Each topic in the vector
is associated with a counter that represents the
number of occurrence of this topic with respect to the
location during a time window t. In the following sub-
sections, modeling of user topics and locations is
explained in details.
3.1.1 Identify User Locations
In order to extract users’ location, we identified all
tweets that contain a foursquare location check-ins.
Then, we utilized the foursquare public API to
identify the location categories
1
for each physical
location extracted from a check-in. We utilized the
Foursquare category hierarchy that consists of two
kinds of nodes, location nodes and category nodes.
A location node represents a distinctive location
such as Starbucks. While, category node represents a
location category such as a coffee. Since members of
social networks usually live in different geographic
locations, they may visit different places belong to the
same category. Accordingly, our proposed
framework relies on distinguishing place category.
Foursquare classification has given us in total 523
location category such as (hotel, stadium, etc…). Due
to the sparsity of available user locations on social
networks such as twitter and Facebook (Gao and Liu,
2014); we used the location category instead of the
actual physical location. Location information is
currently very sparse. Less than 1% of tweets are geo-
Enabling Semantic User Context to Enhance Twitter Location Prediction
225
tagged and information available from the location
fields in users’ profiles is unreliable at best. (Cheng et
al., 2010) Found that only 26% of Twitter users in a
random sample of over 1 million users reported their
city-level location in their profiles and only 0.42% of
the tweets in their dataset were geo-tagged. In order
to overcome this location sparseness problem, we
further classified these location categories into higher
semantic level. We utilized AlchemyAPI
2
(Gangemi,
2013) to further classify all of these location
categories into 23 higher categories that represent the
first tier categories provided by this API such as
(‘sports’, ‘shopping’, ‘food and drink’, etc…).
3.1.2 Identify User Topics
For each tweet that contained a location check-in, we
identified the set of adjacent tweets that have been
posted by the same user within half an hour frame
after the check-in. Then, we extracted the topics from
each of those adjacent tweets using AlchemyAPI,
which resulted in a vector of topics along with their
count of occurrence as shown in Figures 1&2.
Finally, these topics were linked to the parent location
check-in to be used in the upcoming location
prediction component.
Figure 1: Topics Identification and Extraction.
2
www.alchemyapi.com/
Figure 2: Topics Extraction Component.
3.2 Location Prediction
In order to predict current user location category type,
we utilized both tweets’ posting time and user topics
as features for the prediction model. This
classification problem will be explained in detail in
the following subsection.
3.2.1 Features Selection
We used three main features to predict the category
of user current location which are: the topics posted
during half an hour time window stay in a location,
the posting time of the check-in, and the type of the
posting day (either weekend or weekday). In the
following, each of those features is explained in
details.
Topics. The first feature is the vector of topics that
posted by user during her/his stay in specific location
category. Each field of the vector represents the
frequency of occurrence of each topic using
AlchemyAPI topics categorization. In order to
overcome sparsity problem, a threshold variable beta
β is used such that we eliminate this topic from vector
if the frequency of occurrence of a topic is less than
this threshold. This threshold is used to exclude any
noisy topics that may be irrelevant to the location and
hence improve the accuracy of prediction. Finally,
ICAART 2016 - 8th International Conference on Agents and Artificial Intelligence
226
this threshold will also eliminate location check-ins
that have very few topics.
Day Time. According to (Mahmud, Nichols, &
Drews, 2014), user’s posting pattern in twitter
changes over day time. As early morning hours show
less activity than hours in the morning, afternoon, and
evening hours. Therefore, we split day time into 3
categories, morning, working hours and Night. We
assumed the morning check-ins to be any check-in
that has been posted between 12 am to 8 am. The
working hour check-ins will be any check-in posted
from 8 am to 4 pm, and night check-ins will be from
4 pm to 12 am.
Type of Day. A slight shift in the tweeting activity of
the users during weekends is noticed compared to
weekday (Mahmud et al., 2014). Thus, type of day
where the user check-in was considered as third
feature. Type of day is either a weekend if posted on
Saturday or Sunday or it can be a working day if
posted on any other day.
3.2.2 Location Prediction Model
Based on the above three mentioned features
extracted for each check-in we proceed to use
traditional classifiers to predict user current location.
Several classical classifiers are used and applied
using Weka toolkit (Witten and Frank, 2005) such as
Naïve Bayes (NB), C4.5, k-Nearest Neighbor (k-
NN). For KNN different value of K was applied till
10 neighbors with 1/d distance weighting which
provided better accuracy value.
In order to evaluate the accuracy of the proposed
framework we used a baseline probabilistic model
(Cheng et al., 2010). This model relies on the
probability distribution of actual words published by
user in each location. Thus, all words that are related
to a specific location category are extracted and
aggregated from all tweets that have been posted
within half an hour time window after any check-in.
These extracted words are further filtered to remove
any mention tags or stop words.
In this model, prediction of current user’s location
category can be divided into three main steps. First
step is to generate the probability distribution of
words for each location category by calculating the
probability of each word w given the location
category c as shown in equation (1).
(
|
)
=
()
(1)
Where count(w) is the frequency of occurrences of
the word w in all tweets that have been posted within
half an hour timer frame after any location check-in
of type category c. n represent the total number of
words in these tweets. The second step is to calculate
the probability that a user u exist in a specific location
category c based on words ws extracted from his/her
tweets that are posted half an hour after the check-in.
this is done using equation (2).
(
|
)
=
(
|
)
∗
(
)
(2)
Where P(ws|c) represent the total probability of
words ws to exist in location category c. P(ws)
represent the total probability of words ws in the
whole dataset of tweets which represent all textual
tweets extracted for all location categories. Finally,
the probability of set of locations are ranked in order
to select the highest predicted location category based
on the extracted words from half an hour time
window.
4 EXPERIMENT
4.1 Experimental Set Up
Initial dataset that has been used in our previous
research work that represent 1452 public twitter users
with about half million (524,000) tweets (Galal &
ElKorany, 2015). This dataset was prepared for the
following experiments through the following steps.
The first step was the identification of tweets that
contain embedded location check-ins and the second
step was the extraction of the set of adjacent textual
tweets that are posted right after the location check-in
tweets.
4.1.1 Identifying Location Check-in Tweets
In the first step we extracted all tweets that contained
an embedded foursquare location check-ins. This step
gave us in total 16,400 check-in tweets that have been
posted by 1074 users. For each one of these tweets we
extracted its posting time and the embedded check-in
URL.
4.1.2 Extraction of the Set of Adjacent
Textual Tweets
The second step was the extraction of the set of all
adjacent textual tweets for each user that have been
posted within a specific time frame after posting a
location check-in tweet. It is significant to mention
that in order to identify the amount of time which is
considerable enough to post content relevant to the
current place, different time windows have been tried
(an hour and half an hour after check-in). However,
Enabling Semantic User Context to Enhance Twitter Location Prediction
227
prediction accuracy enhanced with half an hour time
frame which is used as our hypothesis that this time
could be considered is the average time for people to
stay in a specific location. Thus, we aggregated
tweets for half an hour after every check-in. This step
gave us a total 30,000 textual tweets.
4.2 Utilizing Semantic of Topics
and Locations Categories
Since location-based social networks suffer from data
sparsity problem, we utilize location categories using
foursquare API. These foursquare locations
categories are further classified using AlchemyAPI
into 23 locations categories. Also we use
AlchemyAPI to identify set of topic of interest per
location category by extracting it from text tweets that
have been posted within half an hour after every
check-in.
4.3 Classification
After the extraction of location categories and the
topics from all aggregated tweets, we started to
perform our classification by using the features to
train our prediction model.
4.3.1 Classification Features
By utilizing the posting time of tweets and the topics
extracted from the adjacent tweets, the features vector
is built as follows:-
1. Posting time of tweets which is either
morning, working hours or night.
2.
The type of the posting day whether it was
a weekday or a weekend.
3.
The frequency of occurrence of each one of
the 23 AlchecmyAPI topics that have
occurred in aggregated tweets.
These features will be used to predict the location
category of the corresponding location check-in
tweet. In order to estimate the correct number of
labeled classes, we collect all location visited by users
and all tweets posted in each location category. Then,
we calculate the coverage of tweets for each location.
Accordingly, and as shown in Figure 3, the top 2
classes (location categories) covers 52 % of the total
check-ins in the dataset while the top 4 and 6 classes
cover 70% and 80% of the total check-ins
respectively. Therefore, we consider only the top
most used 2, 4 and 6 location categories as labeled
classes in the prediction problem. Furthermore, we
compare the prediction accuracy for those classes
with the whole 23-class available in the dataset.
Figure 3: Top-K location Categories Coverage Percentage.
4.3.2 Classification Results
We applied the location prediction using the C4.5
(Mahmud et al., 2014) decision tree algorithm, the
nearest neighbor classifier and Naïve Bayes classifier
with 10 fold cross validation on the four classification
problems. We also used three different values for
threshold beta (which is used to eliminate topics
which are 1 (no threshold), 2 and 3 number of
occurrences) of topic. This threshold is used to
remove any noise topics that have only occurred once
or twice within the half an hour time frame in order
to reduce sparsity problem. Those numbers are used
as further increasing the threshold will lead to empty
training instances that contain all topics with zero
occurrences.
Finally, in order to evaluate our proposed
prediction model we compared it with the baseline
traditional probabilistic classifier which was applied
on words’ probability distribution of tweets.
Figure 4: Comparison between the accuracy of the location
category prediction for each classification problem and
without using threshold for topics count of occurrence.
ICAART 2016 - 8th International Conference on Agents and Artificial Intelligence
228
Figure 5: Comparison between the accuracy of the location
category prediction for each classification problem and
using threshold of minimum two occurrences for any topic.
Figure 6: Comparison between the accuracy of the location
category prediction for each classification problem and
using threshold of minimum three occurrences for any
topic.
By analyzing Table 1 and Figures from 4 to 6, the
accuracy of our context-based prediction model is
extremely high when applied to classification that
discriminates between fewer classes (2 classes). This
coincides with assumptions of researchers who tried
to differentiate between two main locations in users
life time (a user’s home location and work). However,
our proposed location prediction framework also
provided an acceptable accuracy value for 4 and 6
classes representing other locations.
Furthermore, the proposed framework
outperformed the baseline prediction method on all
classification problems especially when increasing
the number of predicted location categories. The
results demonstrates the advantage of utilizing the
semantic of user published content by considering
topics rather than words as well as the significant of
the proposed model to overcome the sparsity of data.
Table 1: Detailed comparison between the different
classification problems using C4.5, NB, 10-NN and the
baseline classifier.
The results showed that the threshold improved
the accuracy of the predication as it removes any
noisy topics that may be irrelevant to the location.
Also a prediction accuracy of 49.8 % is achieved
using threshold β=3 when considering the top-6
location categories that cover 80% of all location
check-ins and an accuracy of 40% % is achieved
when considering all location categories. Also even
without using any threshold an accuracy of 42.8% is
achieved when considering the top-6 location
categories and 33.5% when considering all location
categories in the classification.
5 CONCLUSION
In this paper the semantic relation between topics,
location and time is explored and utilized in a
framework for location category prediction. The
results of the experiment proved the significance of
such relation and how by simple utilization of this
relation can achieve high accuracy in classification
problems for location prediction. Also the experiment
shows the importance of considering the semantic
information rather than terms or words in location
prediction.
In future work we consider enhancing this
framework by utilizing more advanced features such
as user’s friendships and the past history of user’s
check-ins. Also it is considered to use this advanced
version of the framework in prediction of the actual
physical location because majority of users tend to
stay in their city or country for long periods hence
2-Cl ass 4-Cl ass 6-Class 23-Cl ass
10-NN 0.648 0.468 0.425 0.335
Naïve Bayes 0.636 0.452 0.412 0.31
C4.5 0.661 0.467 0.428 0.334
Baseline 0.613 0.373 0.234 0.066
2-Cl ass 4-Cl ass 6-Class 23-Cl ass
10-NN 0.659 0.48 0.449 0.356
Naïve Bayes 0.633 0.449 0.405 0.3
C4.5 0.668 0.49 0.454 0.362
Baseline 0.613 0.373 0.234 0.066
2-Cl ass 4-Cl ass 6-Class 23-Cl ass
10-NN 0.71 0.53 0.498 0.395
Naïve Bayes 0.688 0.501 0.46 0.35
C4.5 0.713 0.534 0.49 0.399
Baseline 0.613 0.373 0.234 0.066
No Threshold
Threshold (β) = 2
Threshold (β) = 3
Enabling Semantic User Context to Enhance Twitter Location Prediction
229
their visited location will not be changed drastically
over short periods especially if we utilized their past
check-ins.
REFERENCES
Abel, F., Gao, Q., Houben, G.-j., & Tao, K. (2011).
Analyzing User Modeling on Twitter for Personalized
News Recommendations. Em User Modeling, Adaption
and Personalization (pp. 1-12).
Bobadilla, J., Ortega, F., Hernando, A., & Gutiérrez, A.
(2013). Recommender systems survey. Knowledge-
Based Systems, 46, 109-132.
Chandra, S., Khan, L., & Muhaya, F. (2011). Estimating
twitter user location using social interactions--a content
based approach. 2011 IEEE Third Int'l Conference on
Privacy, Security, Risk and Trust and 2011 IEEE Third
Int'l Conference on Social Computing (pp. 838-843).
IEEE.
Cheng, Z., Caverlee, J., & Lee, K. (2010). You are where
you tweet: a content-based approach to geo-locating
twitter users. Proceedings of the 19th ACM
international conference on Information and
knowledge management (pp. 759-768). ACM.
Cho, E., Myers, S., & Leskovec, J. (2011). Friendship and
mobility: user movement in location-based social
networks. Proceedings of the 17th ACM SIGKDD
international conference on Knowledge discovery and
data mining (pp. 1082-1090). ACM.
Dalvi, N., Kumar, R., & Pang, B. (2012). Object matching
in tweets with spatial models. Proceedings of the fifth
ACM international conference on Web search and data
mining (pp. 43-52). ACM.
Galal, A., & ElKorany, A. (2015). Dynamic Modeling of
Twitter Users. Proceedings of the 17th International
Conference on Enterprise Information Systems, 2, pp.
585-593. Barcelona, Spain.
Gao, H., & Liu, H. (2014). Data analysis on location-based
social networks. Em Mobile social networking (pp.
165-194). Springer.
Hecht, B., Hong, L., Suh, B., & Chi, E. (2011). Tweets from
Justin Bieber's heart: the dynamics of the location field
in user profiles. Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (pp. 237-
246). ACM.
Jamali, S., & Rangwala, H. (2009). Digging digg: Comment
mining, popularity prediction, and social network
analysis. WISM 2009. International Conference on Web
Information Systems and Mining (pp. 32-38). IEEE.
Kleanthous, S., & Dimitrova, V. (2008). Modelling
Semantic Relationships and Centrality to Facilitate
Community Knowledge Sharing. Em Adaptive
Hypermedia and Adaptive Web-Based Systems (pp.
123-132). Springer Berlin Heidelberg.
Lee, M.-j., & Chung, C.-w. (2011). A User Similarity
Calculation Based on the Location for Social Network
Services. 16th international conference on Database
systems for advanced applications, (pp. 38-52).
Li, N., & Chen, G. (2009). Analysis of a location-based
social network. CSE'09. International Conference on
Computational Science and Engineering. 4, pp. 263-
270. IEEE.
Mahmud, J., Nichols, J., & Drews, C. (2014). Home
location identification of twitter users. ACM
Transactions on Intelligent Systems and Technology
(TIST), 5(3), 47.
Quercia, D., Askham, H., & Crowcroft, J. (2012).
TweetLDA : Supervised Topic Classification and Link
Prediction in Twitter. Proceedings of the 3rd Annual
ACM Web Science Conference, (pp. 247-250).
Ryoo, K., & Moon, S. (2014). Inferring Twitter user
locations with 10 km accuracy. Proceedings of the
companion publication of the 23rd international
conference on World wide web companion (pp. 643-
648). International World Wide Web Conferences
Steering Committee.
Wang, J., & Prabhala, B. (2012). Periodicity based next
place prediction. Nokia Mobile Data Challenge 2012
Workshop. p. Dedicated task, Vol. 2. No. 2. .
Weerkamp, W., & De Rijke, M. (2012). Activity prediction:
A twitter-based exploration. SIGIR Workshop on Time-
aware Information Access.
Witten, I. H., & Frank, E. (2005). Data Mining: Practical
machine learning tools and techniques. Morgan
Kaufmann.
Ye, J., Zhu, Z., & Cheng, H. (2013). What’s your next
move: User activity prediction in location-based social
networks. Proceedings of the SIAM International
Conference on Data Mining. SIAM.
Gangemi, A. (2013) 'A Comparison of Knowledge
Extraction Tools for the Semantic Web', in The
Semantic Web: Semantics and Big Data, Springer
Berlin Heidelberg, pp.351-366.
ICAART 2016 - 8th International Conference on Agents and Artificial Intelligence
230
