the influence of indirect ties on social network dynamics
DESCRIPTION
Social NetworksTRANSCRIPT
The Influence of Indirect Ties on Social Network Dynamics
Xiang Zuo1, Jeremy Blackburn2, Nicolas Kourtellis3,John Skvoretz1 and Adriana Iamnitchi1
1University of South Florida – Florida, USA2Telefonica Research – Barcelona, Spain
3Yahoo Labs – Barcelona, Spain
Network Dynamics
2The spread of cheating behavior in an
online game social network
http://not-ionic.tumblr.com/
What Is an Indirect Tie?
An indirect tie is defined as a relationship between two individuals who have no direct relation but are connected through other node(s) in the network.
3
Why Study Indirect Ties?
4
Indirect ties are known to be a strong forceshaping the network dynamics.
We lack quantitative studies of theinfluence of indirect ties on networkdynamics, especially for social distanceslonger than 2 hops.
Outline
5
Datasets and indirect tie measurements
Indirect ties and link prediction
Timing of link formation
Indirect ties and information diffusion paths
Datasets
6
Networks Nodes Edges APL Edge weights D OT
TF2 2,406 9,720 4.2 [1-21,767] 12 300 days
IE 410 2,765 3.6 [1-191] 9 90 days
CA-I 348 595 6.1 [1-52] 14 N/A
CA-II 1,127 6,690 3.4 [1-127] 11 N/A
Datasets vary from online gaming networks to face-to-face interaction networks and co-authorship networks
Indirect Tie Measurements
7
JaccardCoefficient
(J)
Adamic-Adar (AA)
Social Strength
(SS)
J(a,b) =G(a)Ç G(b)
G(a)È G(b)
AA(a,b) =
1
log G(Z)ZÎG(a)∩G(b)
å
AA(a,b) =1
log(2)+
1
log(3)= 5.42
SSn(a,b) = 1- PpÎRa ,b
n(1-
minc,...kÎN ( p)
[NW (a,c),...NW (k,b)]
n)
NW (a,k) =
w (a,k,l)"lÎLi ,k
å
w (i,u,l)"lÎLi ,u
å"uÎNi
å
Outline
8
Datasets and indirect tie measurements
Indirect ties and link prediction
Timing of link formation
Indirect ties and information diffusion paths
Using Indirect Ties for Link Prediction
9
Can we use a 2-hop (or 3-hop) indirect tie between nodes that are not directly connected to predict whether a link will form between them?
Link Prediction Results
10
Networks n Classifier Metric Precision Recall F-Measure AUC
TF2 2
DecisionTree(J48)
SS 0.75±0.012 0.74±0.008 0.74±0.008 0.77±0.009
AA 0.71±0.004 0.71±0.004 0.71±0.004 0.71±0.006
J 0.51±0.007 0.51±0.006 0.50±0.008 0.51±0.008
IE 2SS 0.84±0.013 0.84±0.002 0.84±0.002 0.87±0.001
AA 0.69±0.002 0.69±0.002 0.68±0.003 0.70±0.003
J 0.69±0.007 0.68±0.005 0.68±0.001 0.68±0.004
TF2 3 SS 0.63±0.020 0.63±0.010 0.62±0.010 0.64±0.030
IE 3 SS 0.64±0.010 0.63±0.010 0.63±0.010 0.66±0.010
Indirect ties are able to predict the formation of links even when the social path is longer than 2.
Outline
11
Datasets and indirect tie measurements
Indirect ties and link prediction
Timing of link formation
Indirect ties and information diffusion paths
Timing of Link Formation
12
Link formation delay: the
interval between the time when
the link formation conditions are
met and the time when the link
forms.
Is there any connection between the strength of an indirect tie and the delay of link formation?
Link Formation Delay Definition
13
Tie Classification
14
Classify indirect ties into strong and weak with three
criteria:
J(a,b) > mean
AA(a,b) > mean
SS(a,b) > mean [C-mean]:
[C-max]:J(a,b) > max
AA(a,b) > max
SS(a,b) > max
J(a,b) > min
AA(a,b) > min
SS(a,b) > min
[C-min]:
Tie (a,b) is a strong
indirect tie
Tie Strength vs. Link Delay
15
2−hop ties with C−max cr iterion in TF2 2−hop ties with C−mean cr iterion in TF2 2−hop ties with C−min cr iterion in TF2
●
●
●●
●
● ●
●
● ●●
●
●●
●●
●
●
0.1
0.2
0.3
0.4
0−1 1−7 7−14 14−30 30−60 60+ 0−1 1−7 7−14 14−30 30−60 60+ 0−1 1−7 7−14 14−30 30−60 60+
Fra
ction ● strong tie
weak tie
3−hop ties with C−max cr iterion in TF2 3−hop ties with C−mean cr iterion in TF2 3−hop ties with C−min cr iterion in TF2
●
●
● ●
● ●
●●
● ● ●
●
●
● ●● ●
●
0.0
0.1
0.2
0.3
0.4
0.5
0−1 1−7 7−14 14−30 30−60 60+ 0−1 1−7 7−14 14−30 30−60 60+ 0−1 1−7 7−14 14−30 30−60 60+
Delay in Days
Fra
ction
2−hop ties with C−max criterion in IE 2−hop ties with C−mean criterion in IE 2−hop ties with C−min criterion in IE●
●● ● ● ●
●
●● ● ● ●
●
●● ● ● ●0.0
0.2
0.4
0.6
0.8
0−5 5−10 10−15 15−20 20−25 25+ 0−5 5−10 10−15 15−20 20−25 25+ 0−5 5−10 10−15 15−20 20−25 25+
Fra
ction
3−hop ties with C−max criterion in IE 3−hop ties with C−mean criterion in IE 3−hop ties with C−min criterion in IE●
●
●
● ● ●
●
●
●●
●
●
0.0
0.2
0.4
0.6
0−5 5−10 10−15 15−20 20−25 25+ 0−5 5−10 10−15 15−20 20−25 25+ 0−5 5−10 10−15 15−20 20−25 25+
Delay in Minutes
Fra
ction
3−hop ties with C−max cr iterion in TF2
●
●
● ●
● ●0.0
0.1
0.2
0.3
0.4
0.5
0−1 1−7 7−14 14−30 30−60 60+
Delay in Days
Fra
ctio
n33% (strong)
vs. 7% (weak)
Strong indirect ties form direct links quicker both in 2 and 3 hops.
Outline
16
Datasets and indirect tie measurements
Indirect ties and link prediction
Timing of link formation
Indirect ties and information diffusion paths
Can Indirect Ties Predict Diffusion Paths?
17
Given that a user received a piece of information at time step t, can we predict which other users will receive this information at time step t+2 or t+3?
Experimental Setup
18
Ground Truth
Linear Threshold (LT) model
in LT
Scale the value of as
ω [1-10] in CA-I
ω [1-30] in CA-II
ω [1-100] in TF2
Path Prediction
Calculate strength of indirect ties
Rank a user’s 2(3)-hop neighbors based on the calculated strength values
Define a cut-off threshold to select the user’s topNindirect neighbors
Î
Î
Î
q
random(0,1) /wq pred
Compare
q = random(0,1)
Prediction Evaluation (2-hop paths)
19
CA−I CA−II TF2
●
●
● ●●
●
●
●●
● ●●
●
●
●●
●
●
0.50.60.70.80.9
2.5 5.0 7.5 10.0 10 20 30 25 50 75 100
Accura
cy
● Adamic−Adar Baseline Jaccard Social Strength
CA−I CA−II TF2●
●
● ●
●
●
●●
●
●
●
●
●
●
●●
●●
0.2
0.4
0.6
0.8
1.0
2.5 5.0 7.5 10.0 10 20 30 25 50 75 100
Sensitiv
ity
CA−I CA−II TF2
●
● ●
●●
●
●
●
●
● ●
●
●●
●
●
●
●
0.4
0.6
0.8
1.0
2.5 5.0 7.5 10.0 10 20 30 25 50 75 100Diffusion Parameter w
Spe
cific
ity
Prediction Evaluation (3-hop paths)
20
CA−I CA−II TF2
●
●
●
● ●
●
●
● ● ●
● ●
●
● ●●
●●
0.5
0.6
0.7
0.8
0.9
2.5 5.0 7.5 10.00 10 20 30 25 50 75 100
Accu
racy
● Baseline Social Strength
CA−I CA−II TF2
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
0.4
0.5
0.6
0.7
0.8
0.9
2.5 5.0 7.5 10.00 10 20 30 25 50 75 100
Se
nsitiv
ity
CA−I CA−II TF2
● ● ● ●
●
●
●
●●
●●
●
●●
●● ●
●
0.4
0.6
0.8
1.0
2.5 5.0 7.5 10.00 10 20 30 25 50 75 100
Diffusion Parameter w
Sp
ecific
ity
Indirect ties can serve as a predictor for information diffusion paths.
Summary
Indirect ties have the power to predict link
formation between people at social distances
greater than 2.
The strength of an indirect tie positively
correlates to the speed at which a direct link
forms between the two people.
Indirect ties can serve as a predictor for
diffusion paths in social networks.21
22
Thanks!
Distributed Systems Grouphttp://www.cse.usf.edu/dsg/