text
stringlengths
0
89.3k
B Adapted EHO for information foraging results
1 Empirical parameters setting
We conducted extensive tests for the sake of tuning the em
pirical parameters to values that ensure the best results in
terms of relevance and response time ie maximizing the
similarity between the users interests and the surfing path
while minimizing the execution time It is important to note
that the stochastic aspect of the EHO algorithm requires to
test each parameter value multiple times to get stable out
comes For that purpose we run the tests at least 100 times
for each parameter
First we started with parameters α01andβ01
Recall that αis a scale parameter that determines the influ
ence of the matriarchs position on rest of the elephants of
the same clan while βdetermines the influence of the aver
age position of the clan on the matriarchs position To select
the best values of both parameters we combined each value
ofαin the range 01with all possible values of βalso in
the same range Figure 7a shows the similarity score results
while Figure7b displays the response time results in seconds
The 3D representation gives a good visualization of the simi
larity and time evolution with the variation of the parameters
With respect to the results showed in both figures we set αto
09andβto04
Another important combination of parameters is the number
of clans and the number of elephants in each clan A proper
number will help to visit different parts of the social graph
and therefore get closer to relevant posts The results shown
in Figure 8a and Figure 8b allow to determine the adequate
number of clans and the number of elephants in each clanThe number of clans is set to 8 with 90elephants in each
clan
The number of generations is the parameter that allows the
algorithm to evolve a sufficient amount of time so it can
reach better results and approach the global optimum We
can observe from Figure 9 that the best number of genera
tions would be 40 since it maximizes the similarity and min
imizes the response time
2 Foraging results
Table 4 presents some examples of the adapted EHO for in
formation foraging results with 7 different users interests
column one generated for evaluation purpose The surf
ing path with the most relevant tweet is displayed in column
two the similarity value between the surfing path and the
users interests is shown in column three and the response
time in seconds alongside the length of the surfing path are
displayed in columns four and five respectively Note that
when the surfing depth is greater than 1 the entire surfing
path is displayed in chronological order of access as in the
case of the users interest diabetes type 2 intermittent fast
ing for example
We observe that in almost all cases the system is capable of
finding relevant tweets However the response time is rela
tively long mainly because of the big size of the social graph
and the fact that the foraging process happens exclusively on
line We can also notice that the surfing depth is to a certain
extent small which can be explained by the fact that the so
cial graph is not strongly connected Moreover during the
construction of the surfing path a tweet is only inserted if it
is more relevant than the tweets that were accessed before it
in the same path
C Enhanced EHO for large scale information foraging re
sults
Although we were able to reach relevant posts using our first
attempt based on the adaptation of the original EHO algo
rithm to information foraging the results showed some lim
itations related to the efficiency especially when it comes to
big social graphs To cope with this issue we proposed in
Section VI a novel approach consisting in an enhanced ver
sion of EHO for large scale infomration foraging
1 Empirical parameters setting
The first parameter to define is the number of territories
ie the number of clusters k For this purpose we tested
the kmeans algorithm with different values of kin the in
terval 180 For each fixed number of clusters k we use
Formula 13 to compute the total Within Cluster Sums of
Squares WSS which measures the average distance be
tween the posts and their corresponding centroids for each
cluster 36 37
WSS kX
i1X
eSide mi 13
Where
k is the number of clusters39 Y Drias et al
a Similarity variation
b Time variation
Figure 7 Setting αandβparameters based on Time and Similarity variation
a Similarity variation
b Time variation
Figure 8 Setting the number of clans and elephants based on Time and Similarity variation
a Similarity variation
b Time variation
Figure 9 Setting the number of generations based on Time and Similarity variationEnhanced Elephant Herding Optimization for Large Scale Information Access on Social Media 40
Main topic Subtopics
Computer science Machine learning Deep learning Artificial intelligence Big data Graph database Open Data IoT 5G Social graph
Cyber security Cyberattack Blockchain Bitcoin Hack IBM Data science Power BI Robotics Smart city Smart
Home Digital Predictive Analytics Mathematics Cisco selfdriving cars VMware Virtual reality Web domains
TensorFlow
Politics American express Free speech Black lives matter Time is up Immigration Twitterstorian Brexit UK V ote Presi
dent Trump Democracy Breaking News Democrats Racism white supremacy
Health Covid19 flatern the curve Vaccine Cholera intermittent fasting Sugar free diet Healthcare HealthTech