Mar 30, 2010 - "Is there any other point to which you would wish to draw my attention?â "To the curious incident of the dog in the night-time.â "The dog did ...
Beyond Degree Distribution:
local to global structure of social contact graphs Stephen Eubank
Chris Barrett, Jiangzhuo Chen, Maleq Khan, Madhav Marathe,
Achla Marathe, Samarth Swarup, Anil Vullikanti NDSSL Tech Report 10-041
Network Dynamics and Simulation Science Lab Virginia Bioinformatics Institute at Virginia Tech March 30, 2010 2010 International Conference on
Social Computing, Behavioral Modeling, and Prediction
Network Dynamics and Simulation Science Laboratory
A multi-method approach to measurement in the social sciences
+ An example application: social network analysis
Network Dynamics and Simulation Science Laboratory
Funded in part by NIGMS Network Dynamics and Simulation Science Laboratory
&
Successful interventions in social systems
depend on identifying causes • A ⇒ B ⇒ C C • A
B
•
Causation is too strong; influence
•
Causation is slippery – A ⇒ B & A precedes B in time – A ⇒ B & (not A) ⇒ (not B)
•
Proximal vs remote cause
•
Causal feedback loops
•
Operational definition:
cause as control surface
•
Deductive analysis: consequences of hypothesized causes
•
A
⇒
C
B
• A
⇒
C B
Network Dynamics and Simulation Science Laboratory
Sherlock Holmesʼs Scientific Method
Doesnʼt Always Work! “When you have eliminated the impossible,
whatever remains, however improbable, must be the truth”
Problems: 1. “Whatever remains” is rarely unique 2. Hypotheses are often not obviously impossible: "Is there any other point to which you would wish to draw my attention?”
"To the curious incident of the dog in the night-time.”
"The dog did nothing in the night-time.”
"That was the curious incident," remarked Sherlock Holmes
Network Dynamics and Simulation Science Laboratory
Constructing a synthetic information system:
choose resolution • Represent each interactor
Network Dynamics and Simulation Science Laboratory
Constructing a synthetic information system:
disaggregate observations • Represent each interactor • Endow them with attributes
Network Dynamics and Simulation Science Laboratory
Constructing a synthetic information system:
estimate interactions • Represent each interactor • Endow them with attributes • Model interaction patterns Contact / colocation network
Network Dynamics and Simulation Science Laboratory
Constructing a synthetic information system:
increase fidelity • Represent each interactor Contact / colocation network • Endow them with attributes • Model interaction patterns • Model other dynamics Communication network
Network Dynamics and Simulation Science Laboratory
Constructing a synthetic information system:
across-level agency • Represent each interactor • Endow them with attributes • Model interaction patterns • Model other dynamics • Unencapsulated agency
Network Dynamics and Simulation Science Laboratory
Synthetic information systems deduce
associations implied by causal hypotheses 1. Network topology affects outcome:
e.g. # infected as a function of time
disease
Generates associations between attribute and probability of infection
Input to contact network model may be remote cause, but network itself is proximal cause
Network Dynamics and Simulation Science Laboratory
Synthetic information systems deduce
associations implied by causal hypotheses 2. Other attributes may just be
“along for the ride”
disease
E.g. Attribute in lower network may be associated with attributes that determine contact network
Network Dynamics and Simulation Science Laboratory
Synthetic information systems deduce
associations implied by causal hypotheses 3. Interactions in one layer may affect network topology in another layer
disease
Associations evolve
information Network Dynamics and Simulation Science Laboratory
Synthetic information systems deduce
associations implied by causal hypotheses 1. Network topology affects outcome:
irreducible complexity 2. Other attributes may just be
“along for the ride”:
generating confounders 3. Interactions in one layer may affect network topology in another layer:
co-evolving networks
Network Dynamics and Simulation Science Laboratory
This approach leads to novel insights The example that follows: • is not a particular question about interventions in a social system • Is a new methodology for social network analysis that just scrapes the surface of whatʼs possible
Network Dynamics and Simulation Science Laboratory
Local network structure affects outcomes How do simple interactions at small scales generate
complex phenomena at large spatio-temporal scales? (micro/macro, genotype/phenotype, individual/population, …)
Individual transmission
Local network structure affects outcomes How do simple interactions at small scales generate
complex phenomena at large spatio-temporal scales? (micro/macro, genotype/phenotype, individual/population, …)
Individual transmission
Simple group transmission
Local network structure affects outcomes How do simple interactions at small scales generate
complex phenomena at large spatio-temporal scales? (micro/macro, genotype/phenotype, individual/population, …)
Individual transmission Courtesy David Nadeau, SDSC
Complex population transmission
Simple group transmission
Local network structure affects outcomes How do simple interactions at small scales generate
complex phenomena at large spatio-temporal scales? (micro/macro, genotype/phenotype, individual/population, …)
The network is as important a part of the dynamics as transmission parameters are. Individual transmission Courtesy David Nadeau, SDSC
Complex population transmission
Simple group transmission
We know social networks exhibit structure …
Comparing observed features of a sociogram with the predictions of a random graph is as old as sociometry itself. Moreno's “sociodynamic effect'' (Moreno, 1934) states that the observed distribution of choices received is more variable than that predicted by chance – there are more “stars" and “isolates" than expected in a random graph.
– P. Holland and S. Leinhardt (eds.), Perspectives on social network research. New York: Academic Press. 1979.
Network Dynamics and Simulation Science Laboratory
… but we donʼt know what it is
As for the taxonomy of large sociograms, this apparently involves problems of great complexity. It would seem offhand that a taxonomy of “nets” … would arise naturally from the consideration of the statistical parameters, e.g. as a continuum of nets in the parameter space. But the statistical parameters themselves are singled out on the basis of taxonomic considerations, which have yet to be clarified.
– Anatol Rapoport and William Horvath,
Behav Sci. 1961, 6, 279–291
Network Dynamics and Simulation Science Laboratory
Consider the space of all graphs
2
There are 2N graphs on N vertices. When the vertices are distinct, many of these graphs are topologically distinct.
Network Dynamics and Simulation Science Laboratory
Some of these graphs have a
particular degree distribution
Scale-free
Proofs about properties of scale-free graphs (e.g. “no epidemic transition in a scale-free graph”) describe mean outcomes across the hexagonal region, sometimes variances.
Network Dynamics and Simulation Science Laboratory
Some of these could represent social networks
Scale-free
Social nets The set of social networks may overlap the set of scale free graphs, but may lie in an out-of-the-way corner. Most of the weight in an average is not from realistic social networks.
Network Dynamics and Simulation Science Laboratory
Degree, Clustering, Assortativity, etc.
are measures of local network structure # vertex
24
edge
26
Deg Deg
#
Degree
#
Clustering
#
1
7
0
21
1
2
2
2
8
2/3
2
1
3
4
3
8
1
1
1
5
1
4
0
2
2
3
5
1
2
3
6
2
5
2
3
3
6
3
5
2
Color
Color
#
Red
Red
1
Red
Yellow
4
Red
Green
7
Red
Blue
3
Yellow Yellow
2
Yellow Green
2
Yellow Blue
2
Green Green
0
Green Blue
3
Blue
2
Blue
Create “more random” graphs
by swapping edge endpoints
Network Dynamics and Simulation Science Laboratory
Constrained edge swaps maintain
local structural invariants
Network Dynamics and Simulation Science Laboratory
Different swapping rules maintain
different structural constraints
This choice maintains degree distribution invariant
Network Dynamics and Simulation Science Laboratory
Different swapping rules maintain
different structural constraints
This choice maintains assortativity by color invariant
Network Dynamics and Simulation Science Laboratory
Different swapping rules maintain
different structural constraints
?
No choice maintains assortativity by degree invariant
Network Dynamics and Simulation Science Laboratory
Different swapping rules maintain
different structural constraints
Any choice maintains clustering distribution invariant
Network Dynamics and Simulation Science Laboratory
Edge swapping is one step in a Markov chain
Markov chain
Our Markov chain mixes quickly, sampling from all graphs with the constrained property. It simulates the entire distribution of outcomes across the space – not just mean and variance – and demonstrates where the starting point lies in that distribution. Network Dynamics and Simulation Science Laboratory
The chainʼs starting point is important
Similar studies have gone the other direction, introducing structure into random graphs. There is no guarantee the result is relevant to social networks.
Network Dynamics and Simulation Science Laboratory
age assortative
infections
degree preserving
days cumulative infections
days
Conclusions • Local structure of interaction networks – beyond degree distribution – has important effects on dynamics in social systems • High resolution representations of social systems suggest innovative techniques, e.g. an edge-swapping Markov Chain for network analysis • Synthetic information systems provide a new methodology to
deduce associations implied by multi-theory, multi-perspective,
causal hypotheses
Network Dynamics and Simulation Science Laboratory