Visualizing the Evolution of Community Structures in Dynamic Social Networks Khairi Reda Department of Computer Science University of Illinois at Chicago.

Visualizing the Evolution of Community Structures in Dynamic Social Networks

Khairi Reda

Department of Computer ScienceUniversity of Illinois at Chicago

Social Networks Analysis (SNA)

• SNA is concerned with structures of ties in the social system, rather than behavior of individual actors

• Visualization has been a central theme in SNA since its inception

• Graphs are the most common visual representation

• Efficient graph layouts make structural patterns emerge while reducing clutter

• Provide a static snapshot of the network - but social systems are dynamic

J. Moreno. Who Shall Survive, 1953

Dynamic networks

Static

Grevy’s zebrascommunities

T1 T2

Time related questions

• How do diseases/information spread through population?

• How do social structures (communities) change with outside circumstances?

• What is the lifespan of a social structure, and are there recurring structures?

Actual

Limitations

• Node movement should be minimized to maintain “mental map” => Potentially poor local layouts

• Limited short-term visual memory

• Can see momentarily changes, but not long-term patterns

• Scalability hampered

Dynamic Networks Vis Animated graphs

Limitations

• Substantial redundancy => limited scalability

• Layout not necessarily stable across time slices

• Suffers form 3D artifacts

T1

T2

T3

Groh et al. Dyson, 2009

Corman et al., 2003

Dynamic Networks Vis Stacked graphs

Design goals• Social scientists need to understand how the social structure

evolves and reacts to external circumstances

• Communities (social groups) are among the most important phenomena

• A group of actors interacting closely and frequently

• Fluid membership: individuals switch community affiliation over time

• Dynamic community = dynamic clusters

• Community = identity

Vis need to show evolution of communities along with domain variables to enable cause-effect analysis

Community identification•Given an interaction sequence, assign a

community color to each individual at every timestep.

•Assumptions

•Individual are reluctant to switch community afffiliation - switching cost

•Individual mostly are seen with their own community - visiting cost

•Individuals are rarely absent from their own community - absence cost

•Minimize the total cost across all individuals over interaction sequence

•Temporal resolution maintained at its finest-grain

Community identification•Given an interaction sequence, assign a

community color to each individual at every timestep.

•Assumptions

•Individual are reluctant to switch community afffiliation - switching cost

•Individual mostly are seen with their own community - visiting cost

•Individuals are rarely absent from their own community - absence cost

•Minimize the total cost across all individuals over interaction sequence

•Temporal resolution maintained at its finest-grain

Movie narrative charts, http://xkcd.com/657/

Visual metaphor

N. Wook Kim et al. TimeNets, AVI ’10

Visual metaphor

QRXY

ABC

Community affiliation

switch

• 500 roll-call votes between Jan 13 to July 30, 2010

• 434 legislators

• Each vote considered to occur in a separate timestep (500 timesteps)

• Individuals casting the same vote (Aye, Nay, or Not Voting) considered to be interacting with each other at that time

• Communities = political opinion groups

Case study: Visualizing communities in the US House of Representatives

Case study: Visualizing communities in the US House of Representatives

March 15 - vote to consider debating Kucinich’s resolution for Afghanistan troops withdrawal by 2010

Actual vote on Kucinich’s resolution

Case study: Visualizing communities in the US House of Representatives

Actual vote on Kucinich’s resolution

Liberal democrats - supporting withdrawal

Republicans and conservative democrats - opposing discussion of proposal

Centrist democrats - opposing discussion of proposal

Main stream democrats - supporting discussion of proposal

Case study: Visualizing communities in the US House of Representatives

March 15 - vote to consider debating Kucinich’s resolution for Afghanistan troops withdrawal by 2010

User study

• Behavioral ecologists want to understand how ecological factors (resources, predation risk, etc.) influence the social structure of group-living populations

• Grevy’s zebras

• Endangered population of about 3,000

• Fission-fusion social structure

• 35 individuals observed over a period of 3 months in 2003 in Laikipia, Kenya

• Social interactions inferred from physical proximity

• Four ecology researchers analyzed their Grevy’s zebra dataset using our visualization.

• Session was video and audio taped followed by a short interview

User studyCommunity

movement in spaceCommunity

timelineIndividual

PurpleCommunity

OrangeCommunity

User study• Community timeline was intuitive to

domain scientists

“This is a very clean depiction of community membership. It is easier to see the individuals move [between communities]”

• Supports correlation of attributes with structural changes in the network

“We are looking at a different project that shows the individual by [reproductive] state moving in and out of the community”

“This says what the males, lactating, and non-lactating females are doing. It is very powerful analysis to see when the switch happens”

Stallion

Lactating Female

Non-Lactating Female

Bachelor

User study• Layout stability

“Once we know this is a community, to see the individuals aligned very consistently like this in almost what looks like a British subway map with simple angles is very useful”

• Integration of community timeline and movement data

“[This visualization] finally put time and space-together. This allows us to understand the physical decision making that lead to the shaping of communities. The dynamic community analysis gave us a better picture for understanding zebra dynamics. The space will give us even a better picture of that temporality”

Limitations

• Scalability

• Visualization scales well with number of individuals and timesteps, but less so with number of communities

• US congress ~ 400 individuals, ~8 communities

• Zebra dataset ~35 individuals, ~12 communities

• Layout optimization

• Minimization of thread crossings, locally and globally

Conclusions

• Social network visualization needs to catch up and propose solutions for dynamic social networks

• Graph layouts have limitations when applied to dynamic networks

• The community structure timeline provides an alternative to stacked graphs

• Shows coherent, fine-grained view of the evolving community structure

• Integration of domain data allow cause-effect analysis

Thank you

Khairi Reda - [email protected]

Electronic Visualization LaboratoryUniversity of Illinois at Chicago

Computational Population Biology LaboratoryUniversity of Illinois at Chicago

Funded in part by the National Science Foundationgrants CNS-0821121 and OCI-0943559

•Ogawa et al. clusters every timestep independently, yet

Dynamic Networks VisOther variants

Ogawa et al. APVIS, 2007