text
stringlengths
0
89.3k
of the information foraging process For instance if a user is
interested in information about health it would be unneces
sary to search posts talking about sport
Consequently and given the above observations we propose
to introduce a new step in the Elephant Herding Optimization
algorithm consisting of dividing the search space into multi
ple regions in order to explore them more efficiently In addi
tion to modeling the concept of territories in nature this will
improve the performance of the algorithm especially when
dealing with large scale problems
There are numerous methods for grouping similar objects
together they can be either supervised or unsupervised de
pending on whether classes of objects already exist or not
Supervised classification considers classes to insert the ob
jects whereas unsupervised classification generates clusters
as outcome Clustering is the process of organizing objects
into groups whose members are similar A cluster is a col
lection of objects which are consistent internally but clearly
dissimilar to the objects belonging to other clusters One of
the main advantages of clustering over supervised classifica
tion is the fact that it doesnt require predefined classes and It
can therefore be performed with data of different sizes with
out the need of a taxonomy or a training set
In this paper we perform the clustering phase using k
means algorithm which is known for its efficiency with
large datasets and its capacity of working with textual data
32 33 It intends to automatically group a set of nobjects
intokclusters so that objects in a same cluster are similar
to one another while objects from different clusters are dis
similar 34 The grouping decision is based on the distance
between the object and each cluster centroid mean in a way
that each object belongs to the cluster with the nearest cen
troid A centeroid serves as a prototype of its corresponding
cluster and is defined as the average of all the objects in that
cluster The number of clusters kcan be either predefined or
userdefined depending on the problem the number of ob
jects and the goal behind clustering
When it comes to textual data clustering the idea consists in
grouping texts or documents in clusters based on their con
tent similarity In order to achieve this goal a proper doc
ument representation method is necessary We use the vec
tor space model VSM to represent each social post eias a
weighted vector of terms ei wi1 wi2 w iTwhere
Tis the set of terms or features that occur at least once in
the social graph G The detailed clustering process using k
means is presented in the following subsectionsEnhanced Elephant Herding Optimization for Large Scale Information Access on Social Media 34
A Text preprocessing
This preprocessing phase refers to the set of actions that are
applied to the social posts in order to achieve a good sta
tistical representation of the whole collection This phase
is performed before defining the weights of the words using
TFIDF and includes the following actions
Tokenization which consists in dividing the text into
individual words
Removing special characters related to social media
platforms links usernames etc
Deleting common words that dont bring any semantic
meaning to the text using a stop words list
Reducing each word to its root using adequate algo
rithms such as Porter Stemming As a result inflected
words will be grouped under their word stem which is
referred to as a term
The result of the textual preprocessing is a collection of posts
that are each represented by a set of significant terms The
following subsection explain how each post is afterwards
converted into a weighted vector of terms
B Feature extraction with TFIDF
The goal of the feature extraction using TFIDF is to create
a mapping of the textual data into vectors of terms This
vector representation of the social posts is grounded on the
term relevance concept The weight associated to each term
should be proportional to its importance so that terms with
high weight values are considered as relevant This method
consists in increasing the weight of a term when it appears
many times in a post and lowering it when it is common in
many posts We can summarize the TFIDF calculation with
the two following steps
1 Term Frequency TF
The term frequency TFtiejestimates the importance of
a term tiin a post ejbased on how often tiappears in ej
The more frequent a term is in a post the more important it
is in its description We use Formula 8 to compute the term
frequency
TFtiej freq ijP
tlejfreq lj8
With freq ijbeing the number of occurrences of term tiin
postej
2 Inverse Document Frequency IDF
Inverse Document Frequency indicates how commonly a
word is used in a collection of documents A term has an
important characterizing power if its frequency is high in a
particular social post and low in the rest of the posts We
estimate the inverse document frequency using Formula 9
IDF ti logE
ni 9
WhereEis the total number of posts in the social graph
niis the number of posts containing the term ti
3 TFIDF
We compute the weigh of a term tiin a post ejusing the
TFIDscore which is obtained by Formula 10 The