Upload
cachet
View
43
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Predicting Emerging Social Conventions in Online Social Networks. Farshad Kooti * Winter Mason † Krishna Gummadi * Meeyoung Cha ‡ MPI-SWS * Stevens Institute of Technology † KAIST ‡. Metric. Imperial. Linguistic conventions. Hello. Hey. Aloha. How’s it going. - PowerPoint PPT Presentation
Citation preview
CIKM 2012
Predicting Emerging Social Conventions in Online Social Networks
Farshad Kooti* Winter Mason†
Krishna Gummadi* Meeyoung Cha‡
MPI-SWS* Stevens Institute of Technology† KAIST‡
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 2
Imperial
Metric
3
Linguistic conventions
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti
Hey
AlohaHow’s it going
Hello
4
The retweeting convention
Quoting another user while citing the original author
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti
Bob Alice
RT @Bob:CIKM startedCIKM started
RT @Bob:CIKM started
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 5
Why retweeting convention?
o Information-sharing channels are explicit in Twitter
o Specific to Twitter: exposures within the community
o Contained in Twitter, hence capturing all usages
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 6
Twitter dataset
o Used near-complete data from 03-2006 to 09-2009- 54 million users- 1.9 billion tweets- 1.7 billion follow links
o Follow links are a snapshot of the network in 2009
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 7
The retweeting variations
o Searched for syntax token @username
o “Adopter” refers to a user using the variation at least once
Variation # of adopters # of retweets
RT 1,836 K 53,221 K
via 751 K 5367 K
Retweeting 50 K 296 K
Retweet 36 K 110 K
HT 8 K 22 K
R/T 5 K 28 K
♻ 3 K 18 K
Total 2,059 K 59,065 K
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 8
Our study of retweeting convention
1. Characterizing the emergence [ICWSM’12, best paper award]
2. Predicting the adoption process[this work, CIKM 2012]
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 9
Defining prediction problem
Suppose we are given a social network with records of users, their interactions, and times of adoptions. However, information about which variation was adopted by user u at time t is hidden. How reliably we can infer that user u has adopted variation v at time t?
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 10
RT or via or ...?
RT @john: tweet
tweet (RT @joe)
via @jane: tweet
2,053 TWEETS1,738 FOLLOWING1,581 FOLLOWERS
Bob
Motivation & ProblemFeatures impacting adoptionPredictive power & results
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 12
Feature categories
Personal
Social Global
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 13
feature: # of followersPersonal
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 14
features
# of exposures
# of adopter friends
Social
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 15
feature: # of adopter friendsSocial
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 16
feature: adoption dateGlobal
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 17
All the considered features
– # of followers and friends, # of posted tweets and URLs, join date, geo-location
– # of exposures, # of adopter friends
– Time of adoption
Global
Social
Personal
Motivation & ProblemFeatures impacting adoptionPredictive power & results
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 19
Measuring the predictive power of features
o We calculate Information Gain (IG) of each feature, which shows the predictive power
o IG: change in entropy (measure of uncertainty) because of the given feature
o IG(Variation, feat.) = H(Variation) - H(Variation|feat.)
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 20
Predictive power of features: results
Rank Feature Type
1 Date Global
2 # of exposures to RT Social
3 # of posted URLs Personal
4 # of exposures to via Social
5 Join date of adopter Personal
6 # of posted tweets Personal
7 # of RT adopter friends Social
Findings:• # of exposures has more predictive
power than # of adopter friends• Geography is not important
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 21
Prediction methodology
o Using different ML classifiers: Bayesian models, boosting, decision trees, etc.– Bagging yields the best result
o Feature selection techniques to find best subset of features (excluded 8 features)
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 22
Prediction accuracyVariation Accuracy Precision Recall
RT 71.2 72.8 68.1via 72.6 52.1 66.6
Retweeting 98.0 43.1 90.5Retweet 98.5 34.3 80.1
HT 99.7 50.5 84.9R/T 99.8 19.0 81.5
recycle icon 99.9 35.9 82.3Weighted average 72.6 65.7 69.8
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 23
Dealing with unbalanced classes
o Problem:– Most of the adoptions (68%) are RT– A simple classifier of always predicting the most
used variation performs goodo Solution:– Take the same number of cases from two groups
(baseline: 50%)
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 24
Prediction accuracy from balanced data
Variation Accuracy Precision RecallRT 61.3 60.7 63.1via 60.7 60.6 60.1
Retweeting 59.1 58.9 61.8Retweet 56.9 56.6 56.6
HT 82.3 82.8 81.5R/T 77.3 77.0 77.2
recycle icon 81.5 83.1 80.2Weighted average 61.0 60.7 61.5
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 25
Stronger definitions
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 26
Summary
o Predicting adoption of social conventionso Investigated impact of various factors
o Global feature trumps social and personal featureso The number of exposures had more predictive
power than number of adopter friendso Using the features from network is not enough
for a prediction with high accuracy
Prediction of Emerging Social Conventions in OSNs- Farshad Kooti 27
Thank you!