Random Graph Models

Network Science Reading Group
      October 31, 2011
Modeling Complex Networks
• Real-world complex networks contain an
  extremely large number of nodes (n)
• Nodes interact in various ways
  – Capture interactions via a graph
  – If two nodes interact, there is an edge between
    them

• Question: How should edges be placed in
  order to model real world complex networks?
Random Graph Models
• Look at three graph models that rely on a
  “random” placement of edges
  – Different initial conditions and probability
    distributions lead to different types of graphs
• Three common models:
  – Erdos-Renyi (Exponential)
  – Watts-Strogatz (Small-World)
  – Scale-Free/Barabasi-Albert (Power-Law
    Distribution)
Erdos-Renyi
• Erdos-Renyi graph: G(n,p)
  – n: number of nodes
  – p: probability of adding an edge between any two
    nodes
• Mechanism: each possible edge in the graph is
  included with probability p
• What happens as n→∞ for various values of
  p?
Phase Transitions
• If p < 1/n, graph contains many small components
• At p = 1/n, a giant component starts to form
• At p = log(n)/n, the graph is almost surely
  connected

• There is a phase transition at 1/n
• Note that expected number of edges at each
  node is (n-1)p
Characteristics of Erdos-Renyi Graphs
• If connected, average distance between two nodes is
  small (small-world)

• Degree distribution is Poisson:



• Clustering coefficient: number of edges between
  neighbors of a node, divided by total number of
  possible edges between those neighbors
   – Erdos-Renyi graphs tend to have small clustering
     coefficients – do not match real world networks (high
     coefficients)

                                    Figure from “Scale-Free Networks” by Barabasi and Bonabeau
Watts-Strogatz (Small World) Model
• An effort to generate small-world networks with high
  clustering coefficients
• Start with regular lattice and rewire each edge with a
  certain probability p




• Small-world and high clustering coefficient, but degree
  distribution does not match real-world networks
                         Figure from “Statistical Mechanics of Complex Networks” by Albert and Barabasi
Scale-Free Networks
• Real world networks display degree
  distributions that have a power-law
  distribution

               P( k )  k 

• These are called power-law or scale-free
  networks
• Previous random graph models do not
  generate scale free networks
Preferential Attachment
• Start with a small group of nodes
• At each time-step, a new node comes in and
  attaches to existing nodes
  – Key point: prefer to attach to nodes that have a
    higher degree

• Can show that this leads to a network that has
  a scale-free distribution
  – Contains hubs that connect to many nodes
Degree Distribution of Scale-Free
           Networks




                   Figure from “Scale-Free Networks” by Barabasi and Bonabeau

Random graph models

  • 1.
    Random Graph Models NetworkScience Reading Group October 31, 2011
  • 2.
    Modeling Complex Networks •Real-world complex networks contain an extremely large number of nodes (n) • Nodes interact in various ways – Capture interactions via a graph – If two nodes interact, there is an edge between them • Question: How should edges be placed in order to model real world complex networks?
  • 3.
    Random Graph Models •Look at three graph models that rely on a “random” placement of edges – Different initial conditions and probability distributions lead to different types of graphs • Three common models: – Erdos-Renyi (Exponential) – Watts-Strogatz (Small-World) – Scale-Free/Barabasi-Albert (Power-Law Distribution)
  • 4.
    Erdos-Renyi • Erdos-Renyi graph:G(n,p) – n: number of nodes – p: probability of adding an edge between any two nodes • Mechanism: each possible edge in the graph is included with probability p • What happens as n→∞ for various values of p?
  • 5.
    Phase Transitions • Ifp < 1/n, graph contains many small components • At p = 1/n, a giant component starts to form • At p = log(n)/n, the graph is almost surely connected • There is a phase transition at 1/n • Note that expected number of edges at each node is (n-1)p
  • 6.
    Characteristics of Erdos-RenyiGraphs • If connected, average distance between two nodes is small (small-world) • Degree distribution is Poisson: • Clustering coefficient: number of edges between neighbors of a node, divided by total number of possible edges between those neighbors – Erdos-Renyi graphs tend to have small clustering coefficients – do not match real world networks (high coefficients) Figure from “Scale-Free Networks” by Barabasi and Bonabeau
  • 7.
    Watts-Strogatz (Small World)Model • An effort to generate small-world networks with high clustering coefficients • Start with regular lattice and rewire each edge with a certain probability p • Small-world and high clustering coefficient, but degree distribution does not match real-world networks Figure from “Statistical Mechanics of Complex Networks” by Albert and Barabasi
  • 8.
    Scale-Free Networks • Realworld networks display degree distributions that have a power-law distribution P( k )  k  • These are called power-law or scale-free networks • Previous random graph models do not generate scale free networks
  • 9.
    Preferential Attachment • Startwith a small group of nodes • At each time-step, a new node comes in and attaches to existing nodes – Key point: prefer to attach to nodes that have a higher degree • Can show that this leads to a network that has a scale-free distribution – Contains hubs that connect to many nodes
  • 10.
    Degree Distribution ofScale-Free Networks Figure from “Scale-Free Networks” by Barabasi and Bonabeau