25
U. Michigan participation in EDIN Lada Adamic, PI E 2.1 fractional immunization of networks E 2.1 time series analysis approach to correlating structure and content, and co- evolving structure E 2.3 role of groups in information diffusion E 2.3 cultural differences in communication structure INAR C

U. Michigan participation in EDIN Lada Adamic, PI

  • Upload
    shalom

  • View
    34

  • Download
    0

Embed Size (px)

DESCRIPTION

U. Michigan participation in EDIN Lada Adamic, PI. E 2.1 fractional immunization of networks E 2.1 time series analysis approach to correlating structure and content, and co-evolving structure E 2.3 role of groups in information diffusion E 2.3 cultural differences in communication structure. - PowerPoint PPT Presentation

Citation preview

Page 1: U. Michigan participation in EDIN Lada Adamic, PI

U. Michigan participation in EDINLada Adamic, PI

• E 2.1 fractional immunization of networks• E 2.1 time series analysis approach to correlating structure

and content, and co-evolving structure • E 2.3 role of groups in information diffusion• E 2.3 cultural differences in communication structure

INARC

Page 2: U. Michigan participation in EDIN Lada Adamic, PI

2

Fractional Immunization in Hospital-transfer Graphs B. Aditya Prakash1, Lada A. Adamic2, Theodore Iwashyna2, Hanghang

Tong3, Christos Faloutsos1

1Carnegie Melon University, 2University of Michigan,3IBM

Page 3: U. Michigan participation in EDIN Lada Adamic, PI

3

hospital setting

• Hospitals harbor highly resistant bacteria• These bacteria can hitch a ride when patients are transferred

from hospital to hospital

communication network setting

• individuals may propagate misinformation or malicious computer viruses

two settings

Page 4: U. Michigan participation in EDIN Lada Adamic, PI

4

one problem

complete immunization is not feasible

• all prior work on immunization on networks assumes complete immunization

our approach: fractional immunization

• allocating resources to nodes reduces their probability of becoming infected

• e.g. allocating r units of resource corresponds to reducing Prob(infection) to

(0.75)r

Page 5: U. Michigan participation in EDIN Lada Adamic, PI

5

Fractional Asymmetric Immunization

• Fractional Effect• Asymmetric Effect

Page 6: U. Michigan participation in EDIN Lada Adamic, PI

6

Fractional Asymmetric Immunization

Fractional Effect [ f(x) = ]• Asymmetric Effect

Edge weakened by half

x5.0

Page 7: U. Michigan participation in EDIN Lada Adamic, PI

7

Fractional Asymmetric Immunization

Fractional Effect [ f(x) = ] Asymmetric EffectOnly incoming edges

x5.0

Page 8: U. Michigan participation in EDIN Lada Adamic, PI

8

Fractional Asymmetric Immunization

• Fractional Effect [ f(x) = ]• Asymmetric Effect

# antidotes = 3

x5.0

Page 9: U. Michigan participation in EDIN Lada Adamic, PI

9

Fractional Asymmetric Immunization

• Fractional Effect [ f(x) = ]• Asymmetric Effect

# antidotes = 3

x5.0

Page 10: U. Michigan participation in EDIN Lada Adamic, PI

10

Fractional Asymmetric Immunization

• Fractional Effect [ f(x) = ]• Asymmetric Effect

# antidotes = 3

x5.0

Page 11: U. Michigan participation in EDIN Lada Adamic, PI

11

Problem Statement

• Hospital-transfer networks – Number of patients transferred

• Given:– The SI model– Directed weighted graph – A total of k antidotes– A weakening function f(x)

• Find: – the ‘best’ distribution which minimizes the “footprint” at some time t

Page 12: U. Michigan participation in EDIN Lada Adamic, PI

Naïve way

• How to estimate the footprint?– Run simulations? – too slow– takes about 3 weeks for graphs of typical size!

Page 13: U. Michigan participation in EDIN Lada Adamic, PI

Our Solution – Main Idea

• The SI model has no threshold– any infection will become an epidemic

• But– can bound the expected number of infected nodes at time t

• Get the distribution which minimizes the bound!

Page 14: U. Michigan participation in EDIN Lada Adamic, PI

Our Solution – Main Idea

• NP-complete!• We give a fast, effective near-optimal algorithm -

GreedyResync– O(km/r + kN)

Page 15: U. Michigan participation in EDIN Lada Adamic, PI

Simulations

Lower is better

Our algorithm, near optimal

US-MEDICARE Hospital Patient Transfer network

Page 16: U. Michigan participation in EDIN Lada Adamic, PI

16

simulation results

Page 17: U. Michigan participation in EDIN Lada Adamic, PI

17

Resource allocation

few ICU beds

many ICU beds

fewer resources

more resources

Page 18: U. Michigan participation in EDIN Lada Adamic, PI

18

fractional immunization: summary

• Targeted resource allocation is 16x more effective than uniform

• Best strategy: heavily concentrate resources at a few particularly important hospitals

• Greedy algorithm is near-optimal

Page 19: U. Michigan participation in EDIN Lada Adamic, PI

Time series analysis of network co-evolution

• Can the evolution of network structure reveal attributes of the content?– imagine that pattern of who communicates with whom is easy to

discern, but acquiring content is costly (paying informant, decrypting, etc.)

– Can the structure suggest when it would be appropriate to • Can the evolution of one network predict how another

network over the same nodes will evolve in the future?

Chun-Yuen Teng, Liuling Gong, Avishay Livne, Lada Adamic

Twitter data

Page 20: U. Michigan participation in EDIN Lada Adamic, PI

contemporaneous correlation between structure and content

predicting the similarity between non-linked pairs using textual and structural variables

correlation between textual and structural features

Page 21: U. Michigan participation in EDIN Lada Adamic, PI

measuring co-evolution

• temporal conductance– degree of unexpectedness– recent and frequent edges, or those that close recent and frequent

paths, are expected

• Second Life data:– low conductance (network is novel) corresponds to lower entropy in

exchanged assets– “free” asset transfer network time series predicts, via temporal

conductance, paid transaction time series

Page 22: U. Michigan participation in EDIN Lada Adamic, PI

The role of groups in information diffusion

• Main findings:– group variables help to explain adoption

• e.g. overlap of groups an individual and previous adopters belong to• group variables are more predictive than # of adopting contacts, etc.

– group structure is predictive of amount of exchange• e.g. higher clustering

David Huffaker, Chun-Yuen Teng, Liuling Gong, Matthew Simmons, Lada Adamic

Page 23: U. Michigan participation in EDIN Lada Adamic, PI

group structure conducive to exchange

low rates of adoption

high ratesof adoption

Page 24: U. Michigan participation in EDIN Lada Adamic, PI

cultural differences in co-evolving communication patterns

• corporate communication– In Asia, individuals use different channels for different contacts

Jiang Yang1, Zhen Wen2, Lada Adamic1, Mark Ackerman1

1U. Michigan, 2IBM

Page 25: U. Michigan participation in EDIN Lada Adamic, PI

cultural differences in sentiment expression