text
stringlengths
0
89.3k
2for all eiSjdo
3 µjµj ei
4end for
5µj1
Sjµj
6mindist
7postid 1
8for all eiSjdo
9 compute the distance dei µjusing Formula 11
10 ifdei µj min distthen
11 postidi
12 end if
13end for
14mjepostid
15Return mj
Algorithm 5 Kmeans for territories definition
Input E posts of the social graph G
Output S set of kclusters
1initialize kcentroids randomly using Algorithm 2
2repeat
3 create kclusters using Algorithm 3
4 update the centroids using Algorithm 4
5until convergence or max itererations
6return S
social graph mthe size of the vectors kthe number of clus
ters and lthe number of iterations Although the complex
ity is linear it can require a significant time especially with
large scale social graphs Nevertheless this will not affect
the information foraging performance since the clustering is
performed offline and only once
VI Enhanced EHO for large scale information
foraging EEHOLSIF
Addressing large scale information foraging can be tricky
and time consuming The worst case complexity of informa
tion foraging corresponds to the case when the social graph
is a complete graph and is estimated to OQp
i1ni with
nbeing the total number of posts in the graph and pthe surf
ing depth 16 In this paper we propose a new bioinspired
approach to information access based on enhanced elephant
herding optimization using the concepts presented in sections
III IV and V We introduce a new enhanced Elephant Herd
ing Optimization variant to improve the performance of the
original algorithm and adapt it to large scale information for
aging Our contribution focuses on several aspects from the
initialization of the algorithm to the clans structure The
main aspects are detailed in the following subsections
A Semantic position assignment
The concept of territories introduced in section V allows to
considerably optimize the foraging process by delimiting the
search area We exploit the clustering results to assign a nuEnhanced Elephant Herding Optimization for Large Scale Information Access on Social Media 36
merical position to each post on the social graph based on
the posts content Following the clustering process each
post of the social graph is given an integer identifier which
will serve as a position in EEHOLSIF algorithm in a way
that posts belonging to the same cluster have neighbor po
sitions These positions are sorted according to the Euclid
ian distance between posts and the centroid of the cluster so
within the same cluster if a position iis less than another po
sition j this means that the post associated with iis closer
to the centroid than the post associated with j For instance
let us consider a social graph with 4000 posts and 3 clus
ters with S1having 1500 posts S2having 1500 posts and
S3having 998 posts The positions in each cluster will be
distributed as follows
S11500S
i1ei
S23001S
i1501ei
S34000S
i3002ei
B Initialization of the algorithm
Territories definition and semantic positions assignment will
play a major role in the initialization phase of EEHOLSIF
Indeed unlike the original version of the algorithm where
the initialization is performed in a complete random way the
clustering and the semantic positions assignment permit to
target the cluster containing the posts that are the closest to
the users interest and then set the initial elephants positions
accordingly The minimal distance between clans distClan
and the maximal distance between elephants of the same clan
distElephant become more representative since the positions
are assigned based on the content of the posts In fact this
ensures that the elephants of the same clan are browsing pots
that have similar content while elephants of different clans
are located on dissimilar posts with regards to their content
This will result in a better distribution of the elephants on the
search space and a better coverage of the potential solutions
In order to initialize the clans we first compute the euclidean
distance between the users interests vector Iand each cen
troid vector mj Then we launch the different clans either
on the cluster with the nearest centroid or on a cluster cho
sen according to a uniform distribution probability For this
purpose we introduce q a random variable uniformly dis
tributed in 01andq001a tunable parameter We
propose the pseudo random proportional rule for choosing a
territory highlighted by Formula 12
ifq q 0then
Pci Sj
1ifjargmin dI m j