Data Mining and Machine Learning:
Fundamental Concepts and Algorithmsdataminingbook.info
Mohammed J. Zaki1 Wagner Meira Jr.2
1Department of Computer ScienceRensselaer Polytechnic Institute, Troy, NY, USA
2Department of Computer ScienceUniversidade Federal de Minas Gerais, Belo Horizonte, Brazil
Chapter 4: Graph Data
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 1 / 48
Graphs
A graph G = (V ,E ) comprises a finite nonempty set V of vertices or nodes, and aset E ⊆ V ×V of edges consisting of unordered pairs of vertices.
The number of nodes in the graph G , given as |V |= n, is called the order of thegraph, and the number of edges in the graph, given as |E |=m, is called the size
of G .
A directed graph or digraph has an edge set E consisting of ordered pairs ofvertices.
A weighted graph consists of a graph together with a weight wij for each edge(vi ,vj) ∈ E .
A graph H = (VH ,EH) is called a subgraph of G = (V ,E ) if VH ⊆ V and EH ⊆ E .
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 2 / 48
Undirected and Directed Graphs
v1 v2
v3 v4 v5 v6
v7 v8
v1 v2
v3 v4 v5 v6
v7 v8
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 3 / 48
Degree Distribution
The degree of a node vi ∈ V is the number of edges incident with it, and isdenoted as d(vi ) or just di .
The degree sequence of a graph is the list of the degrees of the nodes sorted innon-increasing order.
Let Nk denote the number of vertices with degree k . The degree frequency
distribution of a graph is given as
(N0,N1, . . . ,Nt)
where t is the maximum degree for a node in G .
Let X be a random variable denoting the degree of a node. The degree
distribution of a graph gives the probability mass function f for X , given as
(
f (0), f (1), . . . , f (t))
where f (k) = P(X = k) = Nkn
is the probability of a node with degree k .
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 4 / 48
Degree Distribution
v1 v2
v3 v4 v5 v6
v7 v8
The degree sequence of the graph is
(4,4,4,3,2,2,2,1)
Its degree frequency distribution is
(N0,N1,N2,N3,N4) = (0,1,3,1,3)
The degree distribution is given as(
f (0), f (1), f (2), f (3), f (4))
= (0,0.125,0.375,0.125,0.375)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 5 / 48
Path, Distance and Connectedness
A walk in a graph G between nodes x and y is an ordered sequence of vertices, startingat x and ending at y ,
x = v0, v1, . . . , vt−1, vt = y
such that there is an edge between every pair of consecutive vertices, that is,(vi−1,vi ) ∈ E for all i = 1,2, . . . , t. The length of the walk, t, is measured in terms ofhops – the number of edges along the walk.
A path is a walk with distinct vertices (with the exception of the start and end vertices).A path of minimum length between nodes x and y is called a shortest path, and thelength of the shortest path is called the distance between x and y , denoted as d(x ,y). Ifno path exists between the two nodes, the distance is assumed to be d(x ,y) =∞.
Two nodes vi and vj are connected if there exists a path between them. A graph isconnected if there is a path between all pairs of vertices. A connected component, orjust component, of a graph is a maximal connected subgraph.
A directed graph is strongly connected if there is a (directed) path between all orderedpairs of vertices. It is weakly connected if there exists a path between node pairs only byconsidering edges as undirected.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 6 / 48
Adjacency Matrix
A graph G = (V ,E ), with |V |= n vertices, can be represented as an n× n,symmetric binary adjacency matrix, A, defined as
A(i , j) =
{
1 if vi is adjacent to vj
0 otherwise
If the graph is directed, then the adjacency matrix A is not symmetric.
If the graph is weighted, then we obtain an n× n weighted adjacency matrix, A,defined as
A(i , j) =
{
wij if vi is adjacent to vj
0 otherwise
where wij is the weight on edge (vi ,vj) ∈ E .
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 7 / 48
Graphs from Data Matrix
Many datasets that are not in the form of a graph can still be converted into one.
Let D = {x i}ni=1 (with x i ∈R
d), be a dataset. Define a weighted graphG = (V ,E ), with edge weight
wij = sim(x i ,x j)
where sim(x i ,x j) denotes the similarity between points x i and x j .
For instance, using the Gaussian similarity
wij = sim(x i ,x j) = exp
{
−‖x i − x j‖
2
2σ2
}
where σ is the spread parameter.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 8 / 48
Iris Similarity Graph: Gaussian Similarityσ = 1/
√2; edge exists iff wij ≥ 0.777
order: |V |= n= 150; size: |E |=m= 753
uT
uT
uT
uT
uT
uT
uT
uT
uT
uT
uT
uT
uT
uT
uTuT
uT
uT
uTuT
uTuT
uT
uTuT
uT
uT
uT
uT
uTuT
uTuT
uT
uT
uTuT
uT
uTuT
uT
uTuT
uT
uT
uT
uT
uT
uT
uT
rSrS
rSrS rSrS
rSrS
rS
rS
rS
rS
rSrS
rSrSrS
rS
rS
rS
rS
rS
rS
rSrS
rS
rS
rS
rSrS
rSrS
rS
rS
rS
rS
rS
rS
rS
rSrS
rSrSrS
rS
rS
rSrS
rS
rS
bC
bC
bC
bC
bCbC
bCbC
bCbC
bC
bCbC
bC
bC
bC
bC
bC
bC
bC
bC bC
bCbC
bC
bCbC
bCbC
bCbC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bCbC
bC
bC
bC
bCbC
bC
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 9 / 48
Topological Graph Attributes
Graph attributes are local if they apply to only a single node (or an edge), andglobal if they refer to the entire graph.
Degree: The degree of a node vi ∈ G is defined as
di =∑
j
A(i , j)
The corresponding global attribute for the entire graph G is the average degree:
µd =
∑
i di
n
Average Path Length: The average path length is given as
µL =
∑
i
∑
j>i d(vi ,vj)(
n
2
) =2
n(n− 1)
∑
i
∑
j>i
d(vi ,vj)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 10 / 48
Iris Graph: Degree Distribution
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
Degree: k
f(k)
6
13
8
6
5
8 8
13
10
6
9
6
7
6
5
1 1
2
4 4
3
4
5
1
3
0
1
0
1
2
1
0 0 0
1
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 11 / 48
Iris Graph: Path Length Histogram
0 1 2 3 4 5 6 7 8 9 10 110
100
200
300
400
500
600
700
800
900
1000
Path Length: k
Fre
quen
cy
753
1044
831
668
529
330
240
146
90
30 12
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 12 / 48
Eccentricity, Radius and Diameter
The eccentricity of a node vi is the maximum distance from vi to any other nodein the graph:
e(vi ) = maxj
{
d(vi ,vj)}
The radius of a connected graph, denoted r(G ), is the minimum eccentricity ofany node in the graph:
r(G ) =mini
{
e(vi )}
=mini
{
maxj
{
d(vi ,vj)}
}
The diameter, denoted d(G ), is the maximum eccentricity of any vertex in thegraph:
d(G ) =maxi
{
e(vi )}
=maxi,j
{
d(vi ,vj)}
For a disconnected graph, values are computed over the connected components ofthe graph.
The diameter of a graph G is sensitive to outliers. Effective diameter is morerobust; defined as the minimum number of hops for which a large fraction,typically 90%, of all connected pairs of nodes can reach each other.Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 13 / 48
Clustering Coefficient
The clustering coeff icient of a node vi is a measure of the density of edges in theneighborhood of vi .
Let Gi = (Vi ,Ei ) be the subgraph induced by the neighbors of vertex vi . Note thatvi 6∈ Vi , as we assume that G is simple.
Let |Vi |= ni be the number of neighbors of vi , and |Ei |=mi be the number ofedges among the neighbors of vi . The clustering coefficient of vi is defined as
C (vi ) =no. of edges in Gi
maximum number of edges in Gi
=mi(
ni2
) =2 ·mi
ni(ni − 1)
The clustering coeff icient of a graph G is simply the average clustering coefficientover all the nodes, given as
C (G ) =1
n
∑
i
C (vi )
C (vi ) is well defined only for nodes with degree d(vi )≥ 2, thus define C (vi ) = 0 ifdi < 2.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 14 / 48
Transitivity and Efficiency
Transitivity of the graph is defined as
T (G ) =3× no. of triangles in G
no. of connected triples in G
where the subgraph composed of the edges (vi ,vj) and (vi ,vk) is a connected
triple centered at vi , and a connected triple centered at vi that includes (vj ,vk) iscalled a triangle (a complete subgraph of size 3).
The eff iciency for a pair of nodes vi and vj is defined as 1d(vi ,vj )
. If vi and vj are
not connected, then d(vi ,vj) =∞ and the efficiency is 1/∞= 0.The eff iciency of a graph G is the average efficiency over all pairs of nodes,whether connected or not, given as
2
n(n− 1)
∑
i
∑
j>i
1
d(vi ,vj)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 15 / 48
Clustering Coefficient
v1 v2
v3 v4 v5 v6
v7 v8
Subgraph induced by node v4:
v1
v3 v5
v7
The clustering coefficient of v4 is
C(v4) =2(
4
2
) =2
6= 0.33
The clustering coefficient for G is
C(G ) =1
8
(
1
2+
1
3+ 1+
1
3+
1
3
)
=2.5
8= 0.3125
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 16 / 48
Centrality Analysis
A centrality is a function c : V →R, that induces a ranking on V .
Degree Centrality: The simplest notion of centrality is the degree di of a vertexvi – the higher the degree, the more important or central the vertex.
Eccentricity Centrality: Eccentricity centrality is defined as:
c(vi ) =1
e(vi )=
1
maxj {d(vi ,vj)}
The less eccentric a node is, the more central it is.
Closeness Centrality: closeness centrality uses the sum of all the distances torank how central a node is
c(vi ) =1
∑
j d(vi ,vj)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 17 / 48
Betweenness Centrality
The betweenness centrality measures how many shortest paths between all pairs ofvertices include vi . It gives an indication as to the central “monitoring” role playedby vi for various pairs of nodes.
Let ηjk denote the number of shortest paths between vertices vj and vk , and letηjk(vi ) denote the number of such paths that include or contain vi .
The fraction of paths through vi is denoted as
γjk(vi ) =ηjk(vi )
ηjk
The betweenness centrality for a node vi is defined as
c(vi ) =∑
j 6=i
∑
k 6=ik>j
γjk =∑
j 6=i
∑
k 6=ik>j
ηjk(vi )
ηjk
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 18 / 48
Centrality Values
v1 v2
v3 v4 v5 v6
v7 v8
Centrality v1 v2 v3 v4 v5 v6 v7 v8
Degree 4 3 2 4 4 1 2 2
Eccentricity 0.5 0.33 0.33 0.33 0.5 0.25 0.25 0.33e(vi ) 2 3 3 3 2 4 4 3
Closeness 0.100 0.083 0.071 0.091 0.100 0.056 0.067 0.071∑
j d(vi ,vj) 10 12 14 11 10 18 15 14
Betweenness 4.5 6 0 5 6.5 0 0.83 1.17
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 19 / 48
Prestige or Eigenvector Centrality
Let p(u) be a positive real number, called the prestige score for node u. Intuitivelythe more (prestigious) the links that point to a given node, the higher its prestige.
p(v) =∑
u
A(u,v) · p(u)
=∑
u
AT (v ,u) · p(u)
Across all the nodes, we have
p′ =ATp
where p is an n-dimensional prestige vector.By recursive expansion, we see that
pk =ATpk−1 =
(
AT)2
pk−2 = · · ·=(
AT)k
p0
where p0 is the initial prestige vector. It is well known that the vector pk
converges to the dominant eigenvector of AT .
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 20 / 48
Computing Dominant Eigenvector: Power Iteration
The dominant eigenvector of AT
and the corresponding eigenvaluecan be computed using the power
iteration method.
It starts with an initial vector p0,
and in each iteration, it multiplieson the left by AT , and scales theintermediate pk vector by dividingit by the maximum entry pk [i ] inpk to prevent numeric overflow.
The ratio of the maximum entry initeration k to that in k − 1, givenas λ=
pk [i ]
pk−1[i ], yields an estimate
for the eigenvalue.
The iterations continue until thedifference between successiveeigenvector estimates falls belowsome threshold ǫ > 0.
PowerIteration (A, ǫ):1 k← 0 // iteration
2 p0← 1 ∈Rn// initial vector
3 repeat
4 k← k + 1 pk ←ATpk−1
// eigenvector estimate
5 i ← argmaxj{
pk [j ]}
// maximum
value index
6 λ← pk [i ]/pk−1[i ] // eigenvalue
estimate
7 pk ←1
pk [i ]pk // scale vector
8
9 until∥
∥pk −pk−1
∥
∥≤ ǫ
10 p← 1‖pk‖
pk // normalize eigenvector
11 return p,λ
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 21 / 48
Power Iteration for Eigenvector Centrality: Example
v1
v4 v5
v3 v2
A=
0 0 0 1 00 0 1 0 11 0 0 0 00 1 1 0 10 1 0 0 0
AT =
0 0 1 0 00 0 0 1 10 1 0 1 01 0 0 0 00 1 0 1 0
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 22 / 48
Power Method via Scaling
p0 p1 p2 p3
11111
12212
→
0.511
0.51
11.51.50.51.5
→
0.6711
0.331
11.331.330.671.33
→
0.7511
0.51
λ 2 1.5 1.33
p4 p5 p6 p7
11.51.50.751.5
→
0.6711
0.51
11.51.50.671.5
→
0.6711
0.441
11.441.440.671.44
→
0.6911
0.461
11.461.460.691.46
→
0.6811
0.471
1.5 1.5 1.444 1.462
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 23 / 48
Convergence of the Ratio to Dominant Eigenvalue
0 2 4 6 8 10 12 14 161.25
1.50
1.75
2.00
2.25
bc
bc
bc
bc bc bc bc bc bc bc bc bc bc bc bc bc λ= 1.466
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 24 / 48
PageRank
PageRank is based on (normalized) prestige combined with a random jump
assumption. The PageRank of a node v recursively depends on the PageRank ofother nodes that point to it.
Normalized Prestige: Define N as the normalized adjacency matrix
N(u,v) =
{
1od(u) if (u,v) ∈ E
0 if (u,v) 6∈ E
where od(u) is the out-degree of node u.Normalized prestige is given as
p(v) =∑
u
NT (v ,u) · p(u)
Random Jumps: In the random surfing approach, there is a small probability ofjumping from one node to any of the other nodes in the graph. The normalizedadjacency matrix for a fully connected graph is
N r =1
n1n×n
where 1n×n is the n× n matrix of all ones.Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 25 / 48
PageRank: Normalized Prestige and Random Jumps
The PageRank vector is recursively defined as
p′ = (1−α)NTp+αN
Tr p
=(
(1−α)NT +αNTr
)
p
=MTp
α denotes the probability of random jumps. The solution is the dominanteigenvector of M
T , where M = (1−α)N +αN r is the combined normalizedadjacency matrix.
Sink Nodes: If od(u) = 0, then only random jumps from u are allowed. Themodified M matrix is given as
Mu =
{
Mu if od(u)> 01n1Tn if od(u) = 0
where 1n is the n-dimensional vector of all ones.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 26 / 48
Hub and Authority Scores (HITS)
The authority score of a page is analogous to PageRank or prestige, and it depends onhow many “good” pages point to it. The hub score of a page is based on how many“good” pages it points to. In other words, a page with high authority has many hubpages pointing to it, and a page with high hub score points to many pages that havehigh authority.
Let a(u) be the authority score and h(u) the hub score of node u. We have:
a(v) =∑
u
AT (v ,u) · h(u)
h(v) =∑
u
A(v ,u) · a(u)
In matrix notation, we obtain
a′ =A
Th h
′ =Aa
Recursively, we have:
ak =AThk−1 =A
T (Aak−1) = (ATA)ak−1
hk =Aak−1 =A(AThk−1) = (AA
T )hk−1
The authority score converges to the dominant eigenvector of ATA, whereas the hubscore converges to the dominant eigenvector of AAT .Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 27 / 48
Small World Property
Real-world graphs exhibit the small-world property that there is a short pathbetween any pair of nodes. A graph G exhibits small-world behavior if the averagepath length µL scales logarithmically with the number of nodes in the graph, thatis, if
µL ∝ logn
where n is the number of nodes in the graph.
A graph is said to have ultra-small-world property if the average path length ismuch smaller than logn, that is, if µL≪ logn.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 28 / 48
Scale-free Property
In many real-world graphs it has been observed that the empirical degreedistribution f (k) exhibits a scale-free behavior captured by a power-lawrelationship with k , that is, the probability that a node has degree k satisfies thecondition
f (k)∝ k−γ
Taking the logarithm on both sides gives
log f (k) = log(αk−γ)
or log f (k) =−γ logk + logα
which is the equation of a straight line in the log-log plot of k versus f (k), with−γ giving the slope of the line.
A power-law relationship leads to a scale-free or scale invariant behavior becausescaling the argument by some constant c does not change the proportionality.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 29 / 48
Clustering Effect
Real-world graphs often also exhibit a clustering effect, that is, two nodes aremore likely to be connected if they share a common neighbor. The clusteringeffect is captured by a high clustering coefficient for the graph G .
Let C (k) denote the average clustering coefficient for all nodes with degree k ;then the clustering effect also manifests itself as a power-law relationship betweenC (k) and k :
C (k)∝ k−γ
In other words, a log-log plot of k versus C (k) exhibits a straight line behaviorwith negative slope −γ.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 30 / 48
Degree Distribution: Human Protein Interaction Network|V |= n= 9521, |E |=m= 37060
0 1 2 3 4 5 6 7 8
−14
−12
−10
−8
−6
−4
−2
Degree: log2 k
Pro
bab
ility
:log2f(k)
bCbC
bCbC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC
bC bCbC bCbCbCbC
bC bC bCbC bC bC bC bC
bCbCbC bC
bCbC
bC bC bC bC bC bC bCbC
bC
bC bC
bC
bC
bC
bC bC
bC bC
bC
bCbC bCbC
bC bC bCbC bCbCbCbC bC bCbC
bCbC bCbC bCbCbC
bC bC bCbC bCbCbC bCbC
bC bC bC bCbCbC bC bC bC bC bC bC bCbCbC bC bC bC bC bC bC bC bC bC bC bC bC bC
−γ =−2.15
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 31 / 48
Cumulative Degree DistributionF c (k) = 1−F (k) where F (k) is the CDF for f (k)
0 1 2 3 4 5 6 7 8
−14
−12
−10
−8
−6
−4
−2
0
Degree: log2 k
Pro
bab
ility
:log2F
c(k)
bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bCbCbC
−(γ− 1) =−1.85
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 32 / 48
Average Clustering Coefficient
1 2 3 4 5 6 7 8
−8
−6
−4
−2
Degree: log2 k
Ave
rage
Clu
ster
ing
Coeffi
cien
t:log2C(k)
bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bCbCbC bC bC bC bC bC bC
bC bC
bCbC
bCbCbCbCbCbC
bC
bCbCbC bCbC
bC
bC bC
bC
bC bC bCbC
bC
bC
bCbCbC
bCbC
bC
bC
bC
bC
bC bCbCbCbC bC bC bC
bC
bC
bC
bC
bC
bC
bC
bCbCbC bCbC
bCbCbC
bC
bCbCbC bC
bC bCbC
bCbC
bCbC bC
bCbCbC
bC
bC bCbC bC
bC
bCbC
bCbC bC
bC
bCbCbC
bC bC
bCbC bCbCbCbC
bC
bC
−γ =−0.55
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 33 / 48
Erdös–Rényi Random Graph Model
The ER model specifies a collection of graphs G(n,m) with n nodes and m edges,such that each graph G ∈ G has equal probability of being selected:
P(G ) =1
(
M
m
) =
(
M
m
)−1
where M =(
n
2
)
= n(n−1)2
and(
M
m
)
is the number of possible graphs with m edges(with n nodes).
Random Graph Generation: Randomly select two distinct vertices vi ,vj ∈ V ,and add an edge (vi ,vj) to E , provided the edge is not already in the graph G .Repeat the process until exactly m edges have been added to the graph.
Let X be a random variable denoting the degree of a node for G ∈ G. Let pdenote the probability of an edge in G
p =m
M=
m(
n
2
) =2m
n(n− 1)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 34 / 48
Random Graphs: Average Degree
Degree of a node follows a binomial distribution with probability of success p,given as
f (k) = P(X = k) =
(
n− 1
k
)
pk(1− p)n−1−k
since a node can be connected to n− 1 other vertices.
The average degree µd is then given as the expected value of X :
µd = E [X ] = (n− 1)p
The variance of the degree is
σ2d = var(X ) = (n− 1)p(1− p)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 35 / 48
Random Graphs: Degree Distribution
As n→∞ and p→ 0 the expected value and variance of X can be rewritten as
E [X ] = (n− 1)p ≃ np as n→∞
var(X ) = (n− 1)p(1− p) ≃ np as n→∞ and p→ 0
The binomial distribution can be approximated by a Poisson distribution withparameter λ, given as
f (k) =λk
e−λ
k!
where λ= np represents both the expected value and variance of the distribution.
Thus, ER random graphs do not exhibit power law degree distribution.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 36 / 48
Random Graphs: Clustering Coefficient and Diameter
Clustering Coefficient: Consider a node vi with degree k. Since p is the probability of anedge, the expected number of edges mi among the neighbors of a node vi is simply
mi =pk(k − 1)
2
The clustering coefficient is
C(vi ) =2mi
k(k − 1)= p
which implies that C(G ) = 1
n
∑
i C(vi ) = p. Since for sparse graphs we have p → 0, this
means that ER random graphs do not show clustering effect.
Diameter: Expected degree of a node is µd = λ, so in one hop a node can reach λ
nodes. Coarsely, in k hops it can reach λk nodes. Thus, we have
t∑
k=1
λk≤ n, which implies that t = log
λn
It follows that the diameter of the graph is
d(G )∝ logn
Thus, ER random graphs are small-world.Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 37 / 48
Watts–Strogatz Small-world Graph Model
The Watts–Strogatz (WS) model starts with a regular graph of degree 2k , havingn nodes arranged in a circular layout, with each node having edges to its k
neighbors on the right and left.The regular graph has high local clustering. Adding a small amount ofrandomness leads to the emergence of the small-world phenomena.
Watts–Strogatz Regular Graph: n= 8, k = 2v0
v1
v2
v3
v4
v5
v6
v7
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 38 / 48
WS Regular Graph: Clustering Coefficient and Diameter
The clustering coefficient of a node v is given as
C (v) =mv
Mv
=3k − 3
4k − 2
As k increases, the clustering coefficient approaches 34
because C (G ) = C (v)→ 34
as k→∞. The WS regular graph thus has a high clustering coefficient.
The diameter of a regular WS graph is given as
d(G ) =
{
⌈
n2k
⌉
if n is even⌈
n−12k
⌉
if n is odd
The regular graph has a diameter that scales linearly in the number of nodes, andthus it is not small-world.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 39 / 48
Random Perturbation of Regular Graph
Edge Rewiring: For each edge (u,v) in the graph, with probability r , replace v
with another randomly chosen node avoiding loops and duplicate edges.
The WS regular graph has m= kn total edges, so after rewiring, rm of the edgesare random, and (1− r)m are regular.
Edge Shortcuts: Add a few shortcut edges between random pairs of nodes, withr being the probability, per edge, of adding a shortcut edge.
The total number of random shortcut edges added to the network is mr = knr .The total number of edges in the graph is m+mr = (1+ r)m= (1+ r)kn.
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 40 / 48
Watts–Strogatz Graph: Shortcut Edgesn= 20, k = 3
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 41 / 48
Properties of Watts–Strogatz Graphs
Degree Distribution: Let X denote the random variable denoting the number ofshortcuts for each node. Then the probability of a node with j shortcut edges is given as
f (j) = P(X = j) =
(
n′
j
)
pj (1− p)n
′−j
with E [X ] = n′p = 2kr and p = 2krn−2k−1
= 2krn′
.
The expected degree of each node in the network is therefore 2k +E [X ] = 2k(1+ r).The degree distribution is not a power law.
Clustering Coefficient: The clustering coefficient is
C(v)≃3(k − 1)
(1+ r)(4kr + 2(2k − 1))=
3k − 3
4k − 2+ 2r(2kr + 4k − 1)
Thus, for small values of r the clustering coefficient remains high.
Diameter: Small values of shortcut edge probability r are enough to reduce the diameterfrom O(n) to O(logn).
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 42 / 48
Watts-Strogatz Model: Diameter (circles) and Clustering
Coefficient (triangles)
0 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.200
10
20
30
40
50
60
70
80
90
100
Edge probability: r
Dia
met
er:d(G
)
bC
bCbC
bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC
bC 167
uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT uT
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Clu
ster
ing
coeffi
cien
t:C(G
)
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 43 / 48
Barabási–Albert Scale-free Model
The Barabási–Albert (BA) yields a scale-free degree distribution based onpreferential attachment; that is, edges from the new vertex are more likely to linkto nodes with higher degrees.
Let Gt denote the graph at time t, and let nt denote the number of nodes, and mt
the number of edges in Gt .
Initialization: The BA model starts with G0, with each node connected to its leftand right neighbors in a circular layout. Thus m0 = n0.
Growth and Preferential Attachment: The BA model derives a new graphGt+1 from Gt by adding exactly one new node u and adding q ≤ n0 new edgesfrom u to q distinct nodes vj ∈ Gt , where node vj is chosen with probability πt(vj)proportional to its degree in Gt , given as
πt(vj) =dj
∑
vi∈Gtdi
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 44 / 48
Barabási–Albert Graphn0 = 3, q = 2, t = 12
At t = 0, start with 3 vertices v0, v1, andv2 fully connected (shown in gray).
At t = 1, vertex v3 is added, with edgesto v1 and v2, chosen according to thedistribution
π0(vi ) = 1/3 for i = 0,1,2
At t = 2, v4 is added. Nodes v2 and v3
are preferentially chosen according to theprobability distribution
π1(v0) = π1(v3) =2
10= 0.2
π1(v1) = π1(v2) =3
10= 0.3
v0
v1
v2
v3
v4
v5
v6
v7 v8
v9
v10
v11
v12
v13
v14
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 45 / 48
Properties of the BA Graphs
Degree Distribution: The degree distribution for BA graphs is given as
f (k) =(q+ 2)(q+ 1)q
(k + 2)(k + 1)k·
2
(q+ 2)=
2q(q+ 1)
k(k + 1)(k + 2)
For constant q and large k, the degree distribution scales as
f (k)∝ k−3
The BA model yields a power-law degree distribution with γ = 3, especially for largedegrees.
Diameter: The diameter of BA graphs scales as
d(Gt) =O
(
logntlog lognt
)
suggesting that they exhibit ultra-small-world behavior, when q > 1.
Clustering Coefficient: The expected clustering coefficient of the BA graphs scales as
E [C(Gt)] =O
(
(lognt)2
nt
)
which is only slightly better than for random graphs.Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 46 / 48
Barabási–Albert Model: Degree Distributionn0 = 3, t = 997,q = 3
1 2 3 4 5 6 7
−14
−12
−10
−8
−6
−4
−2
Degree: log2 k
Pro
bab
ility
:log2f(k)
bC
bCbC
bCbC bC bC
bCbCbC bCbC bCbC bC bCbCbC bC bCbCbC bC bC bC bC
bC bCbCbC
bC
bCbCbC bC bCbCbCbC
bC
bC
bC
bC
bCbCbC
bC
bCbCbC bCbCbC
bC
bC
bC
bC
bC
bC bC
bC bCbCbC
bC
bCbC bC
bC
bC bC
bC bC bC bC bC bC bC bC
−γ =−2.64
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 47 / 48
Data Mining and Machine Learning:
Fundamental Concepts and Algorithmsdataminingbook.info
Mohammed J. Zaki1 Wagner Meira Jr.2
1Department of Computer ScienceRensselaer Polytechnic Institute, Troy, NY, USA
2Department of Computer ScienceUniversidade Federal de Minas Gerais, Belo Horizonte, Brazil
Chapter 4: Graph Data
Zaki & Meira Jr. (RPI and UFMG) Data Mining and Machine Learning Chapter 4: Graph Data 48 / 48