analysis of contagions in multi-layer and multiplex ...oyagan/talks/ccs_multiplex.pdfanalysis of...

49
Analysis of Contagions in Multi-layer and Multiplex Networks Osman Ya˘ gan Department of ECE Carnegie Mellon University Joint work with Yong Zhuang and Alex Arenas Supported by NSF CCF # 1422165 Osman Ya˘ gan (CMU) PhysPlex , Sept. 21, 2016 1 / 49

Upload: others

Post on 06-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Analysis of Contagions in Multi-layerand Multiplex Networks

Osman Yagan

Department of ECE

Carnegie Mellon University

Joint work with Yong Zhuang and Alex Arenas

Supported by NSF CCF # 1422165

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 1 / 49

Page 2: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Dynamical processes on complex networks

∗ Spreading of an initially localized effect throughout the whole (or, a verylarge part of the) network.

Diffusion of information, ideas, rumors, fads, etc.

Disease contagion in human and animal populations.

Cascade of failures, avalanches, sand piles.

Spread of computer viruses or worms on the Web.

Flows of data, materials, biochemicals.

Network traffic, congestion.

∗ Barrat et al. Dynamical Processes on Complex Networks, 2008

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 2 / 49

Page 3: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

What is new?

∗ Most research focus on the limited case of a single and non-interactingnetwork.∗ Yet, many real-world systems do interact with each other.

� Social networks are coupled together:Facebook, Twitter, Google+, YouTube, etc.

Q: Dynamical processes on interacting networks?

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 3 / 49

Page 4: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

TODAY

Information Propagation (Simple Contagion)

A music video, a news article, etc.Letting someone know of somethingReceiving one copy is enough (disease-like propagation)

• Influence Propagation (Complex contagion)

• Joining a riot, adopting a behavior, buying a product, etc.• Requires social reinforcement• Having 100 friends joining a riot will be different than having

only one who does so.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 4 / 49

Page 5: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Simple Contagions/Information Propagation

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 5 / 49

Page 6: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Information Propagation in Multi-layer Networks

Dynamics of information spreading changed dramatically with theonline social networks; e.g., Facebook, Twitter, Instagram, etc.

A key observation: Social networks are coupled with each other andwith the physical communication network

due to people who participate in multiple networks

Q: How does the coupling affect the speed and extent of informationpropagation?

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 6 / 49

Page 7: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 7 / 49

Page 8: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Issues with existing approaches

• Single-layer:

Individuals engage in different types of networks;e.g., Facebook, Whatsapp, Twitter, etc.

• No-clustering:

Social network has a propensity that two friends of one individualare more likely to know each other.

There is lack of study taking these two factors into consideration.

Our approach: Clustered multi-layer/multiplex networks.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 8 / 49

Page 9: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

An Illustration of a Multi-layer/Multiplex Network Model

Figure: An illustration of a multi-layer/multiplex network model.

Each layer is generated randomly from given degree distributions withtunable clustering coefficient.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 9 / 49

Page 10: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Random Networks with Clustering (Newman and Miller)

Clustering: Informally defined as the propensity of two neighbors of a nodeto be neighbors as well.

Friend of a friend is usually a friend

Formal definitions: Strongly related with the number of triangles

Cglobal =3× (number of triangles in network)

number of connected triples, (1)

Clocal =1

n∗

∑i

Ci , (2)

where

Ci =number of triangles connected to vertex i

number of connected triples centered on vertex i, (3)

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 10 / 49

Page 11: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

A simple example

(a) (b)

(a): global coef. = 0.2; local coef. = 0.3

(b): global coef. = 0.4; local coef. = 0.7

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 11 / 49

Page 12: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Clustered Random Networks

Generated randomly from given degree distributions with tunableclustering.

Configuration Model:

Stub types: only single stubs.The degree of a node: d .Degree distribution: pd

Generalized Configuration Model:

Stub types: single stubs and triangle stubs.The degree of a node: d = (ds , dt)Total degree of a node: ds + 2× dt .Degree distribution: pds ,dt

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 12 / 49

Page 13: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Modeling Information Propagation via SIR Epidemics

Susceptible (S): Not aware of the information

Infectious (I): Has the information and currently spreading it

Recovered (R): No longer spreading the information

Tij = 1− e−rijτi : probability that an infectious individual i transmits theinformation to a susceptible contact jrij : rate of contact over the link from i to j ; τi is recovery time for i

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 13 / 49

Page 14: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

SIR Process → Bond Percolation in Multi-layer networks

Assume for simplicity two layers

W : Physical NetworkF : Facebook

Different transmissibility over W and F: Twij and T f

ij

Average transmissibility in W (with i.i.d. rwij and τw1 = · · · = τwn )

Tw := 〈Twij 〉 = 1−

∫ ∞0

e−rτwPw (r)dr .

Bond Percolation Model: Each edge in W (resp. F) is occupiedwith probability Tw (resp. Tf ) independently from all others.

Information propagates only through occupied edges

SIR Model is isomorphic to bond percolation processOsman Yagan (CMU) PhysPlex , Sept. 21, 2016 14 / 49

Page 15: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Problem Formulation and Quantities of Interest

The propagation is triggered by infecting an arbitrary node andcontinues according to SIR model.

final outbreak size: # of nodes that eventually receive the information

self-limited outbreak: outbreak sizen → 0 as n→∞

epidemic: outbreak sizen → e > 0 as n→∞

epidemic threshold : critical boundary in the space of all networkparameters that distinguishes

P[epidemic] > 0 and P[epidemic] = 0

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 15 / 49

Page 16: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Main Results

• Y. Zhuang and O. Yagan, “Information propagation in clusteredmultilayer networks,” IEEE Trans. Network Science and Engineering, 2016

i) the epidemic boundary;

ii) the relative final size of epidemics when P[epidemic] > 0

iii) the exact probability P[epidemic] in the super-critical regime

Main Techniques

Map SIR process to bond percolation

Explore the emergence of a giant component in the resultingmulti-layer network through a multi-type branching process

Use generating functions approach to analyze the branching process

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 16 / 49

Page 17: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Experiments

α : fraction of people who use Facebook

Tw = Tf

0.2 0.4 0.6 0.8 1

Epidem

icSize

0

0.2

0.4

0.6

0.8 α = 0.1− Exp

α = 0.1− Thm

α = 0.5− Exp

α = 0.5− Thm

α = 0.9− Exp

α = 0.9− Thm

(a) Poisson distribution.

Tw = Tf

0 0.2 0.4 0.6 0.8 1

Epidem

icSize

0

0.2

0.4

0.6

0.8

1

α = 0.1− Exp

α = 0.1− Thm

α = 0.5− Exp

α = 0.5− Thm

α = 0.9− Exp

α = 0.9− Thm

(b) Power-law distribution.

Perfect agreement between analysis and simulations!

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 17 / 49

Page 18: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Impact of clustering on information propagation

We use the following degree distributions for number of single edges andtriangles in two networks.

Network F Network W

Single Edges Distribution Poi(2λf ) 2 Poi( 4−c2 λw )

Triangle Edges Distribution Poi(λf ) Poi(c2λw

)Table: Parameters of the doubly Poisson distribution.

c ↑ ⇒ clustering ↑

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 18 / 49

Page 19: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Impact of clustering on epidemic size

c

0 1 2 3 4

Epidem

icSize

0

0.02

0.04

0.06

0.08

0.1α = 0.2

α = 0.3

α = 0.4

Figure: Poisson distribution

c ↑ ⇒ clustering ↑ ⇒ epidemic size ↓

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 19 / 49

Page 20: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Impact of clustering on epidemic boundary

Tf

0 0.1 0.2 0.3 0.4

TW

0

0.1

0.2

0.3

0.4

c = 0.01,α = 0.1

c = 2.00,α = 0.1

c = 3.99,α = 0.1

c = 0.01,α = 0.9

c = 2.00,α = 0.9

c = 3.99,α = 0.9

Figure: critical boundary of epidemics

c ↑ ⇒ clustering ↑ ⇒ Critical Threshold ↑

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 20 / 49

Page 21: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Lesson Learned

Clustering has an inhibitive effect on epidemics

Can be attributed to the fact that the edges used for completingwedges to triangles is redundant for information propagation

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 21 / 49

Page 22: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

An interesting question

Figure: Which case facilitates the propagation of information??

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 22 / 49

Page 23: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

A multi-faceted picture

α ↑ ⇒ large but loose→

{epidemic threshold ↑epidemic size ↑

More difficult to trigger an information epidemic in large online socialnetwork that is loosely connected vs. small but densely connected.

But, when information transmissibility is high, final epidemic size islarger with a large but loosely connected online social network

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 23 / 49

Page 24: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Complex Contagions

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 24 / 49

Page 25: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

A dynamical process: Binary decisions with externalities

• Each individual must decide between two actions, e.g.,� To buy or not to buy a smart phone� To vote for Democrats or Republicans

Simple Threshold Model (D. Watts, PNAS, 2002)

• Nodes can be in either one of the two states: active or inactive.

• Each node is initially given a threshold τ drawn from Pth(τ).

• An inactive node with m active neighbors and k −m inactiveneighbors will turn active if m

k ≥ τ .

• Assumption: Once active, a node can not be deactivated.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 25 / 49

Page 26: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

An Illustration of the Simple Threshold Model

τ = 0.2

(a) mk = 1

6 < τ = 0.2: Fail to be influenced.

(b) mk = 5

6 > τ = 0.2: Influenced.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 26 / 49

Page 27: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Global cascades

• Start by activating a few nodes (give incentives, free samples)

• Global Cascades: A positive fraction of nodes (in the asymptoticlimit) eventually becomes active

∗ Condition, Probability, and Expected size of global cascades when anarbitrary (set of) node(s) is made active?

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 27 / 49

Page 28: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Issues with existing approaches

• Simplex: the networks with only one single link type.

Individuals engage in different types of relationships; e.g.,family, friends, office-mates, college-mates, etc.

• No-clustering:

Social network has a propensity that two friends of one individualare more likely to know each other.

There is lack of study taking these two factors into consideration.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 28 / 49

Page 29: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Our approach

A content-dependent threshold model of contagion(Yagan & Gligor, Physical Review E, 2012)

Clustered (random) multiplex networks

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 29 / 49

Page 30: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

A content-dependent threshold model for multiplexnetworks (PRE, 2012)

A multiplex network: r types of links.

Content-dependent weights c1, . . . , cr :

ci quantifying the relative bias a type-i link has

An inactive node with degree k :

mi active neighborski −mi inactive neighbors via type-i links∑r

i=1 ki = k

Perceived ratio of active neighbors :=c1m1 + c2m2 + . . .+ crmr

c1k1 + c2k2 + . . .+ crkr≥ τ

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 30 / 49

Page 31: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

An Illustration of the Content-dependent Threshold Model

τ = 0.2.

(a) mk = 1

6 < τ = 0.2: Fail to be influenced.

(b) mk = 3×1

3×1+1×5 > 0.2: Influenced.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 31 / 49

Page 32: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

An Illustration of a Multiplex Network Model

Figure: An illustration of a multiplex network model.

Each link-type is generated randomly from given degree distributionswith tunable clustering coefficient.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 32 / 49

Page 33: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Analysis with random clustered multiplex networks

∗ Let r = 2; i.e., assume that there are only two link types.∗ F: type-1 s single links and t triangles: {pbs,t}.∗ W: type-2 s single links and t triangles: {prs,t}.∗ H = F ∪W with colored distribution {pk}

pk = pbk1,s ,k1,t· prk2,s ,k2,t

, k = (k1,s , k1,t , k2,s , k2,t)

Q: Condition, probability, expected size of global cascades?

Y. Zhuang, A. Arenas, and O. Yagan, “Clustering determines the dynamicsof complex contagions in multiplex networks,” submitted, August 2016.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 33 / 49

Page 34: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Simulation results

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 34 / 49

Page 35: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Agreement between the Analysis and Simulations

pbst = e−λb,1(λb,1)s

s!e−λb,2

(λb,2)t

t!, s, t = 1, 2, . . .

prst = e−λr,1(λr ,1)s

s!e−λr,2

(λr ,2)t

t!, s, t = 1, 2, . . .

(pbs,t): s single links and t triangles in type-1.

(prs,t): s single links and t triangles in type-2.

c = c1/c2 : relative importance of type-1 vs. type-2 links

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 35 / 49

Page 36: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Agreement between Analysis and Simulations (Cont’d)

Degree Parameter (λr,1 = λr,2 = λb,1 = λb,2)0 0.5 1 1.5 2 2.5 3

Fractional

Size

0

0.2

0.4

0.6

0.8

1S - AnlysS - Expt.Ptrigger - AnlysPtrigger - Expt.

Content Parameter (c)0 1 2 3 4 5 6

Fractional

Size

0

0.1

0.2

0.3

0.4

S - Expt.S - Anlys.Ptrigger - Anlys

Figure: Simulations for doubly Poisson degree distributions. In (a), we set thecontent parameter c = 0.25, the threshold as τ = 0.18, and α = 0.5, then varythe degree parameters. In (b), we fix τ = 0.18, λr ,1 = λr ,2 = λb,1 = λb,2 = 0.3,and α = 0.5 while varying content parameter c .

• There is a perfect agreement between our analysis and simulations.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 36 / 49

Page 37: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Impact of Clustering on Influence Propagation

We use the following degree distributions for number of single edges andtriangles in two networks.

Network F Network W

Single Edges Distribution Poi(2λ) 2 Poi( 4−η2 λ)

Triangle Edges Distribution Poi(λ) Poi(η

2λ)

Table: Parameters of the doubly Poisson distribution.

η ↑ ⇒ clustering ↑

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 37 / 49

Page 38: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Impact of Clustering on Influence Propagation (Cont’d)

Degree Parameter (λ)0 0.5 1 1.5 2

FractionalSize

0

0.2

0.4

0.6η = 3.99η = 3η = 0.01

(a) Probability

Degree Parameter (λ)0 0.5 1 1.5 2

ClusteringCoefficient

0

0.1

0.2

0.3

0.4η = 3.99η = 3η = 0.01

(b) Clustering

Degree Parameter (λ)0 0.5 1 1.5 2

FractionalSize

0

0.2

0.4

0.6

0.8

1η = 3.99η = 3η = 0.01

(c) Expected Size

• η ↑ ⇒ clustering ↑.• clustering ↓ Prob. to trigger cascades (low degrees).

• clustering ↑ Prob. to trigger a cascades (high degrees).

• clustering ↓ Expected cascade size (low degrees).

• clustering ↑ Expected cascade size (high degrees).

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 38 / 49

Page 39: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Simplex Networks vs. Multiplex Networks

We focus on the case of

Non-clustered networks

Standard threshold model, i.e., c = 1

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 39 / 49

Page 40: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Degree Distribution

The degree distributions of red and blue edges.

pbk = e−λbλkbk!, k = 0, . . . , (4)

prk = αe−λrλkrk!

+ (1− α)1[k = 0], k = 0, . . . .

where

α: the relative size of nodes that have red and blue edges to that onlyhave blue edges.

λr and λb: the degree parameter of red and blue edges respectively.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 40 / 49

Page 41: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Simplex Projected Theory vs. Multiplex Theory

multiplex theory:

Constraint: red stubs are connected to red ones. blue stubs areconnected to blue ones.

projected theory:

No Constraint: ignores the color of the edges and matches allstubs randomly with each other.

assortativity: the Pearson correlation coefficient of degree between pairsof linked nodes

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 41 / 49

Page 42: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Simplex Projected Theory vs. Multiplex Theory

multiplex theory:

Constraint: red stubs are connected to red ones. blue stubs areconnected to blue ones.high assortativity when α is small.

projected theory:

No Constraint: ignores the color of the edges and matches allstubs randomly with each other.the assortativity is negligible.

assortativity: the Pearson correlation coefficient of degree between pairsof linked nodes

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 42 / 49

Page 43: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Multiplex Networks with Limited Assortativity

We enforce

λr = λb. (5)

projected theory:

the assortativity is negligible.

multiplex theory:

the assortativity is negligible when α is large

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 43 / 49

Page 44: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Multiplex Networks with Limited Assortativity (Cont’d)

Degree Parameter (λr = λb)0 2 4 6 8

Fractional

Size

0

0.2

0.4

0.6

0.8

1

projected theory: α = 0.99multiplex theory: α = 0.99simulations: α = 0.99

Degree Parameter (λr = λb)0 2 4 6 8

Fractional

Size

0

0.2

0.4

0.6

0.8

1

projected theory: α = 0.1multiplex theory: α = 0.1simulations: α = 0.1

projected theory:

α = 0.99: negligible assortativityα = 0.1: negligible assortativity

multiplex theory:

α = 0.99: negligible assortativityα = 0.1: assortativity ≈ 0.21

Slight difference when α = 0.1

Probable suspect: assortativity .

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 44 / 49

Page 45: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Multiplex Networks with Assortativity

To better observe the impact of assortativity, we set

αλr = λb. (6)

If α = 0.1, then λr is 10 times as large as λb.

projected theory:

the assortativity is still negligible.

multiplex theory:

significant assortativity.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 45 / 49

Page 46: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Multiplex Networks with Assortativity (Cont’d)

Degree Parameter (αλr = λb)0 2 4 6 8

Fractional

Size

0

0.2

0.4

0.6

0.8

1

projected theory: α = 0.99multiplex theory: α = 0.99simulations: α = 0.99

Degree Parameter (αλr = λb)0 2 4 6 8

Fractional

Size

0

0.2

0.4

0.6

0.8

1

projected theory: α = 0.1multiplex theory: α = 0.1simulations: α = 0.1

projected theory:

α = 0.99: negligible assortativityα = 0.1: negligible assortativity

multiplex theory:

α = 0.99: negligible assortativityα = 0.1: assortativity ∈ [0.19, 0.79].

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 46 / 49

Page 47: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Two vs. Four Phase Transitions

Degree Parameter (αλr = λb)0 2 4 6 8

Fractional

Size

0

0.2

0.4

0.6

0.8

1

multiplex theory: α = 0.1multiplex theory: α = 0.166multiplex theory: α = 0.99

Figure: The demonstration of multiplex transition phases.

α = 0.1: Cascade only in nodes with red edge.α = 0.166: Cascade in nodes regardless of colors.α = 0.99: Cascade only in nodes with blue edge.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 47 / 49

Page 48: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

Conclusions

We analyze the diffusion of influence in clustered multiplex networks.

We solve analytically for the probability of global cascades andexpected cascade size.

We demonstrate how clustering affects the probability of triggering aglobal cascade and the expected cascade size.

We also make a comparison between influence propagation in simplexand multiplex networks.

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 48 / 49

Page 49: Analysis of Contagions in Multi-layer and Multiplex ...oyagan/Talks/CCS_Multiplex.pdfAnalysis of Contagions in Multi-layer and Multiplex Networks Osman Ya gan Department of ECE Carnegie

THANKS

See www.ece.cmu.edu/~oyagan for references

Osman Yagan (CMU) PhysPlex , Sept. 21, 2016 49 / 49