Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Pinpointing Influence in
Panagiotis Liakos1 Katia Papakonstantinopoulou1 Michael Sioutis2
Konstantinos Tsakalozos3 Alex Delis1
1University of Athens – 2Universite d’Artois, CRIL – 3Canonical Group Ltd.
2nd International Workshop on Social Influence Analysis(co-located with IJCAI 2016)New York City, July 9th, 2016
Motivation
The extreme growth of online social networks enables us tostudy influence patterns at scale.
We want to answer if there exist certain individuals with thepower to affect their social contacts and convince them to buy aproduct or adopt a political idea.
Identifying influential individuals
allows for cost-effective viral marketing techniques to increase brandawareness or even sway the public opinion!
Studies on Twitter [CHBG10] reveal that topological measures suchas indegree fail to capture the influential strength of users.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 2/20
Contribution
We perform an in-depth empirical analysis onand seek to answer:
Is the finding of [CHBG10] true across other online socialnetworks as well, and to what extent?
Does the use of PageRank [LSMW98] allow for a betterestimation of a user’s influential power?
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 3/20
What is Pinterest?
is a visual bookmarking tool that helps youdiscover and save creative ideas.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 4/20
Pinterest lets you:
Pin something to a boardand come back to it later to learn more.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 5/20
Pinterest lets you:
Follow people whose taste you admireto receive their pins in your home feed.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 6/20
Pinterest lets you:
Repin or Like pins of others.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 7/20
Pinterest lets you:
See how others interact with your pins.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 8/20
Why Pinterest?
stands out for many reasons:
It was the fastest site to surpass 10,000,000 monthly active users.
It has more than 100,000,000 monthly active users.
Its vast majority of users are female.
has attracted significant commercial attention:
Users tend to create digital shopping lists of products they areinterested in buying.Therefore, businesses invest in creating compelling boards into increase their revenue.
Michael Sioutis Pinpointing Influence in Pinterest-• Motivation 9/20
Definition of Influence in Pinterest
Indegree influence: the number of followers of a user directlyindicates the size of the audience of that user.
PageRank influence: the PageRank of a user indicates thestrength of her influence on her followers.
Like influence: the number of likes containing one’s nameindicates the ability of that user to generate popular content.
Repin influence: the number of repins containing one’s nameindicates the ability of that user to generate content withpass-along value.
Michael Sioutis Pinpointing Influence in Pinterest-• Our Approach 10/20
Experimental Setting
Dataset [ZSS+14, ZSSS13]:– 36,198,633 users– 983,520,986 social ties– 18,957,340 repins– 9,066,973 likes
PageRank Execution:– Dell PowerEdge R630 server with an Intel R© Xeon R© E5-2630 v3,
2.40 GHz processor, and 256 GB of RAM– Deployed an Apache Hadoop 2.7.1 cluster using Juju1
– Run PageRank as an Apache Giraph process
1https://jujucharms.com/big-data
Michael Sioutis Pinpointing Influence in Pinterest-• Our Approach 11/20
Distribution of indegree and received repins/likes
1
10
100
1000
10000
100000
1e+06
1e+07
1 10 100 1000 10000
Nu
mb
er o
f U
sers
Indegree & Number of Repins/Likes
Indegree, Repin & Like Distributions
IndegreeRepins
Likes
there are a few users with morethan 1,000 followers, repins or likes
most of the activity is centered
around a small minority of users
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 12/20
Distribution of indegree and received repins/likes
1
10
100
1000
10000
100000
1e+06
1e+07
1 10 100 1000 10000
Nu
mb
er o
f U
sers
Indegree & Number of Repins/Likes
Indegree, Repin & Like Distributions
IndegreeRepins
Likes
there are a few users with morethan 1,000 followers, repins or likes
most of the activity is centered
around a small minority of users
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 12/20
Distribution of indegree and received repins/likes
1
10
100
1000
10000
100000
1e+06
1e+07
1 10 100 1000 10000
Nu
mb
er o
f U
sers
Indegree & Number of Repins/Likes
Indegree, Repin & Like Distributions
IndegreeRepins
Likes
there are a few users with morethan 1,000 followers, repins or likes
most of the activity is centered
around a small minority of users
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 12/20
Overlap of Top-Ranked Users
9,955
3,751 3,749
119
6,215
Indegree
Repins Likes
25
9,990
3,757 3,758
23
6,235
PageRank
Repins Likes
5
the overlap of indegree with
both repins and likes is marginal
the overlap of PageRank withrepins and likes is also insignificant
Hints of very weak correlationof the indegree or PageRank of users
with the frequency they receive repins and likes.
Overlap of top-10,000 users ranked by the measures of influenceunder consideration
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 13/20
Overlap of Top-Ranked Users
9,955
3,751 3,749
119
6,215
Indegree
Repins Likes
25
9,990
3,757 3,758
23
6,235
PageRank
Repins Likes
5
the overlap of indegree with
both repins and likes is marginal
the overlap of PageRank withrepins and likes is also insignificant
Hints of very weak correlationof the indegree or PageRank of users
with the frequency they receive repins and likes.
Overlap of top-10,000 users ranked by the measures of influenceunder consideration
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 13/20
Overlap of Top-Ranked Users
9,955
3,751 3,749
119
6,215
Indegree
Repins Likes
25
9,990
3,757 3,758
23
6,235
PageRank
Repins Likes
5
the overlap of indegree with
both repins and likes is marginal
the overlap of PageRank withrepins and likes is also insignificant
Hints of very weak correlationof the indegree or PageRank of users
with the frequency they receive repins and likes.
Overlap of top-10,000 users ranked by the measures of influenceunder consideration
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 13/20
Overlap of Top-Ranked Users
9,955
3,751 3,749
119
6,215
Indegree
Repins Likes
25
9,990
3,757 3,758
23
6,235
PageRank
Repins Likes
5
the overlap of indegree with
both repins and likes is marginal
the overlap of PageRank withrepins and likes is also insignificant
Hints of very weak correlationof the indegree or PageRank of users
with the frequency they receive repins and likes.
Overlap of top-10,000 users ranked by the measures of influenceunder consideration
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 13/20
Comparing Influence Measures
For all measures of influence:
– We assigned the rank of 1 to the most influential user, andincreased the rank as we proceeded to less influential users.
– Identical values were each assigned fractional ranks equal to theaverage of the positions in the ascending order of the values.
We used Spearman’s rank correlation coefficient ρ to examinewhether two rankings covary.
ρ = 1− 6∑d2i
n(n2 − 1)
where di = rg(Xi)− rg(Yi) is the difference between the two ranksof user i, and n is the total number of users.
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 14/20
Rank correlation for all users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
both indegree & PageRank exhibit very
weak correlation with repins and likesassociation is much weaker on Pin-terest than on Twitter [CHBG10]
correlation between repinsand likes is extremely strong
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 15/20
Rank correlation for all users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
both indegree & PageRank exhibit very
weak correlation with repins and likes
association is much weaker on Pin-terest than on Twitter [CHBG10]
correlation between repinsand likes is extremely strong
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 15/20
Rank correlation for all users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
both indegree & PageRank exhibit very
weak correlation with repins and likes
association is much weaker on Pin-terest than on Twitter [CHBG10]
correlation between repinsand likes is extremely strong
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 15/20
Rank correlation for all users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
both indegree & PageRank exhibit very
weak correlation with repins and likesassociation is much weaker on Pin-terest than on Twitter [CHBG10]
correlation between repinsand likes is extremely strong
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 15/20
Rank correlation for the top 10th percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
correlation of indegree with repins
or likes is indeed even weaker for
the top 10th percentile of usersassociation of PageRank withuser influence is about twiceas strong as that of indegree
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 16/20
Rank correlation for the top 10th percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
correlation of indegree with repins
or likes is indeed even weaker for
the top 10th percentile of users
association of PageRank withuser influence is about twiceas strong as that of indegree
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 16/20
Rank correlation for the top 10th percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
correlation of indegree with repins
or likes is indeed even weaker for
the top 10th percentile of users
association of PageRank withuser influence is about twiceas strong as that of indegree
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 16/20
Rank correlation for the top 1st percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
association of PageRank withuser influence is more than twiceas strong than that of indegree
Using PageRank instead of indegree allows forcapturing the importance of users more accurately.
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 17/20
Rank correlation for the top 1st percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
association of PageRank withuser influence is more than twiceas strong than that of indegree
Using PageRank instead of indegree allows forcapturing the importance of users more accurately.
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 17/20
Rank correlation for the top 1st percentile of users
0
0.2
0.4
0.6
0.8
1S
pea
rman
's r
ank
corr
elat
ion
co
effici
ent
Indegree/RepinsPageRank/Repins
Indegree/LikesPageRank/Likes
Repins/Likes
association of PageRank withuser influence is more than twiceas strong than that of indegree
Using PageRank instead of indegree allows forcapturing the importance of users more accurately.
Michael Sioutis Pinpointing Influence in Pinterest-• Empirical Analysis 17/20
References
[CHBG10] Meeyoung Cha, Hamed Haddadi, Fabrıcio Benevenuto, and P. Krishna Gummadi, Measuring UserInfluence in Twitter: The Million Follower Fallacy, ICWSM, 2010.
[LSMW98] Page Lawrence, Brin Sergey, Rajeev Motwani, and Terry Winograd, The PageRank Citation Ranking:Bringing Order to the Web, Technical report, Stanford University, 1998.
[ZSS+14] Changtao Zhong, Mostafa Salehi, Sunil Shah, Marius Cobzarenco, Nishanth Sastry, and Meeyoung Cha,Social Bootstrapping: How Pinterest and Last.fm Social Communities Benefit by Borrowing Links fromFacebook, WWW, 2014.
[ZSSS13] Changtao Zhong, Sunil Shah, Karthik Sundaravadivelan, and Nishanth Sastry, Sharing the Loves:Understanding the How and Why of Online Content Curation, ICWSM, 2013.
Michael Sioutis Pinpointing Influence in Pinterest-• References 18/20
Conclusion - Future Work
The study of influence patterns is essential for the design ofsuccessful advertising strategies.
We performed an in-depth analysis of .
We found that there is very little correlation between the rankingof users based on their indegree and their ranking based on thenumber of either repins or likes they receive.
We proposed the use of PageRank instead of indegree for theidentification of influential users.
We found that Pagerank’s correlation with the ranking of usersbased on repins or likes received is limited, however, muchstronger than that of indegree.
Michael Sioutis Pinpointing Influence in Pinterest-• Conclusion 19/20
thank you!
for further details visit:http://hive.di.uoa.gr/network-analysis/
or email me at: [email protected]
Michael Sioutis Pinpointing Influence in Pinterest-• Conclusion 20/20