Professional Documents
Culture Documents
Problem
Sampling networks always involves the act of aggregation (e.g., when collecting longitudinal samples of networks). We sutdy how the cumulation window length eects the properties of the aggregated network.
Network characteristics are extremely sensitive to minor changes in aggregation length. In our previous work [1] [2], we studied the cumulative properties of Elementary Dynamic Network models over the complete time period (i.e., until they reach the stable point of a full network). Here we focus on the more realistc domain of sparse (cumulative) networks. We nd that even when snapshot networks are stationary, important network characteristics (average path lenght, clustering, betwenness centrality) are extremely sensitive to aggregation (window length).
12 0.30 6 10 0.8 0.25 0 6 20000 40000 clustering 0.20 0.0 0 0.15 0.4 4 0 avgPathLength 8 2
20000
40000
200
400 Time
600
800
1000
0.00 0
0.05
0.10
200
400 Time
600
800
1000
0.06
Basic Concepts
0.5
In our work the dynamic network is a series of graphs, that is, DN = Gt (Vt , Et ), where Et Vt Vt (t 0). The initial network, G0 , is considered as a parameter of the process. The node set xed and we worked with an about constant number of edges. We assume that the evolution of the network can be described as the result of an edge creation and an edge deletion process. We dene Gt as the snapshot network and
T T
maxBetweenness
0.4
0.0 0
0.3
0.6
avgBetweenness
20000
40000
0.04
0.02
0.00
200
400 Time
600
800
1000
0.0 0
0.1
0.2
0.3
20000
40000
200
400 Time
600
800
1000
GT = (
t=0
Vt ,
t=0
Et )
for
T 0.
CPA snapshot
CPA cumulative
SPA snapshot
SPA cumulative
ER1 snapshot
ER1 cumulative
ER2 snapshot
ER2 cumulative
ER3 snapshot
ER3 cumulative
Models
ER1
ing edges. ER3 G0 is a random graph. Rewire kRW edges. SPA (Snapshot preferential ) G0 is a scale free network. Add kA edges from a random node with preferential attachment based on the snapshot network. Delete kD existing edges. CPA (Cumulative preferential ) G0 is a scale free network. Add kA edges from a random node with preferential attachment based on the cumulative network. Delete kD existing edges.
G0 is a random graph. Add each nonexisting edge with pA , delete each existing edge with pD probability. ER2 G0 is a random graph. Add kA uniformly selected random new edges and delete kD exist-
Degree distributions are exceptionally sensitive to the length of the aggregation window. The same
dynamic network may produce a normal, lognormal or even power law distribution for dierent aggregation lenghts. The digree distribution of the snapshot and cumulative network
is inherently dierent. The following surfaces show the CPA model until it approaches the complete network.
CPA snapshot degrees CPA cumulative degrees
100 60
50
80
40 60
p(k)
p(k)
30
40 20 20
10 0
References
[1] [2] Laszlo Gulyas, George Kampis
Ti
k
80
10000
k
80
10000
Eects of Sample Duration on Network Statistics in Elementary Models of Dynamic Networks, International Conference
on Computational Science, Singapore (2011)
Taking slices of the cumulative 3D charts shows us how the degree distribution changes. The log-log charts below show the progression of these changes as the aggregation window gets larger.
ER1 Degree Distribution Comparison, Cumulative Network
1 day 1 year 5 years 10 years 15 years 500
Susan Khor,
100
at the 19th International Symposium on Mathematical Theory of Networks and Systems (MTNS 2010)
p(k)
50
20
Acknowledgements
naNets, FET-Open project no.
p(k)
10
ian Government (KMOP-1.1.2-08/1-2008-0002 ) and the European Union's Seventh Framework Programme: DyFET-233847 (
http:
5 k
10
20
50
10
20
50
100
200
[3]
200
5 k
10
Ti
20
50
//www.dynanets.org).
knowledged.