CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Mining Social-Network Graphs

Lecture 8

Mining Social-Network Graphs

December 4, 2017 1
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Outline
1. Social Networks as Graphs
a) Examples
b) Representation
c) Properties

2. Clustering
a) Distance Measures
b) Girvan-Newman

3. Spectral Analysis
a) Networks as Matrices
b) Connectivity
c) Partitioning
d) Clustering

Mining Social-Network Graphs

December 4, 2017 2
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Social Networks (obvious)

Mining Social-Network Graphs

December 4, 2017 3
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Social Networks (obvious)

Mining Social-Network Graphs

December 4, 2017 4
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Less Obvious
• Phone/E-mail (within K units of time)
• Collaboration (academic papers, patents)
• Biological (proteins, genes)

Mining Social-Network Graphs

December 4, 2017 5
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Representation
• Nodes = entities
– People, papers, …
– Single vs. multiple
• del.icio.us: people, websites, tags

• Edges = relationship
– Discrete vs continuous (i.e. weighted)
– Directed (e.g. following) vs undirected

Mining Social-Network Graphs

December 4, 2017 6
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• Given a graph of 7 nodes (A-G), how
many undirected edges are possible?

Mining Social-Network Graphs

December 4, 2017 7
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Answer
• Given a graph of 7 nodes (A-G), how
many undirected edges are possible?
✓ ◆
7 7! 7(7 1)
= = = 21
2 2!(7 2)! 2

Mining Social-Network Graphs

December 4, 2017 8
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Locality
• Typically assumed that if node A is
connected to both B and C, then it is
more likely than random that B and C are
connected

• Our focus: Community Detection

– Finding groups of densely connected nodes
– (Tightly related to the locality assumption –
i.e. would not work well on random graphs)
Mining Social-Network Graphs

December 4, 2017 9
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example Graph
A B D E

G F

Mining Social-Network Graphs

December 4, 2017 10
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Communities to Be Detected
A B D E

G F

Mining Social-Network Graphs

December 4, 2017 11
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Nodes, Edges, Average Connectivity?

A B D E

G F

• Nodes: 7
• Edges: 9
• Avg Connectivity: 9/21 ~ 0.43

Mining Social-Network Graphs

December 4, 2017 12
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Given A-B, A-C; Expected B-C?

A B D E

G F

• Nodes: 7
• Edges: 9
• Avg Connectivity: 9/21 ~ 0.43
• Expected Locality: 7/19 ~ 0.37

Mining Social-Network Graphs

December 4, 2017 13
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Actual?
A B D E

G F

• Nodes: 7
• Edges: 9
• Avg Connectivity: 9/21 ~ 0.43
• Expected Locality: 7/19 ~ 0.37
• What are all the triples?

Mining Social-Network Graphs

December 4, 2017 14
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Evaluate Node Triples

• A
– BC ✅ A B D E
• B
– AC ✅, AD ❌
– CD ❌ C
• C
G F
– AB ✅
• D
– BE ❌, BF ❌, BG ❌ • F
– EF ✅, EG ❌ – DE ✅, DG ✅
– FG ✅ – EG ❌
• E • G
– DF ✅ – DF ✅

• ✅ = 9, ❌ = 7
• ✅ / (✅ + ❌) = 9/16 ~ 0.56

Mining Social-Network Graphs

December 4, 2017 15
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Local!
A B D E

G F

• Nodes: 7
• Edges: 9
• Avg Connectivity: 9/21 ~ 0.43
• Expected Locality: 7/19 ~ 0.37 Let’s Cluster!
• Actual Locality: 9/16 ~ 0.56 (>> Expected!)

Mining Social-Network Graphs

December 4, 2017 16
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Distance Measure?
• Distance measures on social-network graphs
can be tricky

• KISS ex: D(x,y)=0 if edge, 1 otherwise

– D(A, B)=0
– D(A, C)=0 A B
– D(B, C) = 1 > D(A, B) + D(A, C)

• Also leads to unfortunate ties that C

lead to unfortunate results in
typical clustering algorithms

Mining Social-Network Graphs

December 4, 2017 17
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Betweenness
• One of the simplest measures, based on
finding the edges that are the least likely
to be inside a community
– Clustering: remove high betweenness first!

• B(a, b) = for all pairs of nodes (x, y),

fraction of shortest paths that include
edge ab

Mining Social-Network Graphs

December 4, 2017 18
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Betweenness Intuition (1)

Mining Social-Network Graphs

December 4, 2017 19
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Betweenness Intuition (2)

5 12 4.5
A B D E

1 5
4.5 4 1.5
C

G F
1.5

Mining Social-Network Graphs

December 4, 2017 20
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Betweenness Intuition (3)

A B D E

1
1.5
C

G F
1.5

Mining Social-Network Graphs

December 4, 2017 21
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Computing Graph Betweenness

1. For each node
a) Breadth-First Search (BFS)
• WHY?
b) For each edge, compute betweenness

2. For each edge

a) Sum contributions from trees above
b) Divide by 2 (prevents duplication)

Mining Social-Network Graphs

December 4, 2017 22
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, BFS.0)

E A B D E

G F

Mining Social-Network Graphs

December 4, 2017 23
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, BFS.1)

E A B D E

D F C

G F

Mining Social-Network Graphs

December 4, 2017 24
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, BFS.2)

E A B D E

D F C

G F
B G

Mining Social-Network Graphs

December 4, 2017 25
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, BFS.3)

E A B D E

D F C

G F
B G

A C

Mining Social-Network Graphs

December 4, 2017 26
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, BFS.Done)

E A B D E

D F C

G F
B G

A C

Mining Social-Network Graphs

December 4, 2017 27
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Computing BFS Betweenness

• Recursive credit definition
– Node gets 1 + sum of edge credits below
– Edge gets node below divided by outgoing
• Ignore between nodes at the same level
(never used for shortest paths)

• Start from leaves, move up

Mining Social-Network Graphs

December 4, 2017 28
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.1)

E A B D E

D F C

G F
B G 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 29
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.2)

E A B D E

D F C
0.5 0.5
G F
B G 1
1 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 30
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.3)

E A B D E

D F 1.5 C
0.5 0.5
G F
3 B G 1
1 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 31
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.4)

E A B D E
1.5

D F 1.5 C
3 0.5 0.5
G F
3 B G 1
1 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 32
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.5)

E A B D E
1.5

4.5 D F 1.5 C
3 0.5 0.5
G F
3 B G 1
1 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 33
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (E, Credits.6)

E A B D E
4.5 1.5

4.5 D F 1.5 C
3 0.5 0.5
G F
3 B G 1
1 1

1 A C 1

Mining Social-Network Graphs

December 4, 2017 34
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (A, Credits)

A A B D E
5 1

5 B C 1 C
4
4 G F
D
1 1
1
1 E F G 1
1

Mining Social-Network Graphs

December 4, 2017 35
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (B, Credits)

B A B D E
1 4
1
1
1 A C D 4 C
1 1
1 1 G F
1 E F G 1

Mining Social-Network Graphs

December 4, 2017 36
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (C, Credits)

C A B D E
5 1

5 B A 1 C
4
4 G F
D
1 1
1
1 E F G 1
1

Mining Social-Network Graphs

December 4, 2017 37
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (D, Credits)

D A B D E
1 3
1
1
E F G B 3 C
1 1 1 1
1 G F
1 A C 1

Mining Social-Network Graphs

December 4, 2017 38
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (F, Credits)

F A B D E
1 1
4
4
1 E D G 1 C
3
G F
B 3
1 1

A C
1 1

Mining Social-Network Graphs

December 4, 2017 39
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (G, Credits)

G A B D E
1.5 4.5

1.5 F D 4.5 C
0.5
0.5 3
G F
1 E B 3

1 1
1 A C 1

Mining Social-Network Graphs

December 4, 2017 40
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (A)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B
C
D
E
F
G

Mining Social-Network Graphs

December 4, 2017 41
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (B)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C
D
E
F
G

Mining Social-Network Graphs

December 4, 2017 42
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (C)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D
E
F
G

Mining Social-Network Graphs

December 4, 2017 43
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (D)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E
F
G

Mining Social-Network Graphs

December 4, 2017 44
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (E)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E 1 1 3 4.5 0.5 1.5 0.5
F
G

Mining Social-Network Graphs

December 4, 2017 45
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (F)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E 1 1 3 4.5 0.5 1.5 0.5
F 1 1 3 4 1 1
G

Mining Social-Network Graphs

December 4, 2017 46
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (G)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E 1 1 3 4.5 0.5 1.5 0.5
F 1 1 3 4 1 1
G 1 1 3 0.5 4.5 0.5 1.5

Mining Social-Network Graphs

December 4, 2017 47
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (+)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E 1 1 3 4.5 0.5 1.5 0.5
F 1 1 3 4 1 1
G 1 1 3 0.5 4.5 0.5 1.5
+ 10 2 10 24 9 9 8 3 3

Mining Social-Network Graphs

December 4, 2017 48
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Sum Contributions (/2)

AB AC BC BD DE DG DF EF GF
A 5 1 4 1 1 1
B 1 1 4 1 1 1
C 1 5 4 1 1 1
D 1 1 3 1 1 1
E 1 1 3 4.5 0.5 1.5 0.5
F 1 1 3 4 1 1
G 1 1 3 0.5 4.5 0.5 1.5
+ 10 2 10 24 9 9 8 3 3
/2 5 1 5 12 4.5 4.5 4 1.5 1.5

Mining Social-Network Graphs

December 4, 2017 49
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Edge Contributions
AB AC BC BD DE DG DF EF GF
5 1 5 12 4.5 4.5 4 1.5 1.5

5 12 4.5
A B D E

1 5
4.5 4 1.5
C

G F
1.5

Mining Social-Network Graphs

December 4, 2017 50
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Girvan-Newman
1. Repeat until no edges left
a) Calculate betweenness of edges
b) Remove edge(s) with highest betweenness

Mining Social-Network Graphs

December 4, 2017 51
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (0)

Mining Social-Network Graphs

December 4, 2017 52
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (1)
Step 1: Step 2:

Step 3: Hierarchical network decomposition:

Mining Social-Network Graphs

December 4, 2017 53
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

How to Select # of Clusters?

• Similar to agglomerative, with a different
metric: Modularity (Q; [-1, 1])

• Idea: compare fraction of edges within a

group to the fraction that would be
observed for random connections

Mining Social-Network Graphs

December 4, 2017 54
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Selecting # of Clusters

Mining Social-Network Graphs

December 4, 2017 55
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Direct Partitioning
• We now look at an approach to divide a
graph into two disjoint groups, the task of
bi-partitioning, via spectral analysis

• To do so, we must express the graph as a

matrix

Mining Social-Network Graphs

December 4, 2017 56
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Adjacency Matrix (A)

Aij =1 if (i,j) is an edge, else 0

A B D E

C
2 3
0 1 1 0 0 0 0
61 0 1 1 0 0 07
6 7 G F
61 1 0 0 0 0 07
6 7
60 1 0 0 1 1 17
6 7
60 0 0 1 0 1 07
6 7
40 0 0 1 1 0 15
0 0 0 1 0 1 0
Mining Social-Network Graphs

December 4, 2017 57
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Degree Matrix (D)

Dii = deg(i), else 0

A B D E

C
2 3
2 0 0 0 0 0 0
60 3 0 0 0 0 07
6 7 G F
60 0 2 0 0 0 07
6 7
60 0 0 4 0 0 07
6 7
60 0 0 0 2 0 07
6 7
40 0 0 0 0 3 05
0 0 0 0 0 0 2
Mining Social-Network Graphs

December 4, 2017 58
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Laplacian (L)
L=D-A

A B D E

2 3
2 1 1 0 0 0 0
6 1 3 1 1 0 0 07 G F
6 7
6 1 1 2 0 0 0 07
6 7
6 0 1 0 4 1 1 17
6 7
6 0 0 0 1 2 1 07
6 7
4 0 0 0 1 1 3 15
0 0 0 1 0 1 2

Mining Social-Network Graphs

December 4, 2017 59
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• Rows sum to… ?
0 (degree – edges)

• Columns sum to… ? 2 3

2 1 1 0 0 0 0
0 (degree – edges) 6
6 1 3 1 1 0 0 077
6 1 1 2 0 0 0 07
6 7
• Symmetric? 6
6 0 1 0 4 1 1 17
7
Yes! (D is diagonal, A is 6
6 0 0 0 1 2 1 077
symmetric for undirected 4 0 0 0 1 1 3 15
graph) 0 0 0 1 0 1 2

• Diagonal Dominant?
Yes! (edge case = lonely)

Mining Social-Network Graphs

December 4, 2017 60
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

The Plan
• We will now analyze the eigen-
decomposition (Lv = 𝜆v) of the Laplacian

• It turns out that the smallest eigen-

values/vectors can tell us a lot!
– Note: L is PSD, so 𝜆 ≥ 0
– 0: connectivity
– Next smallest: partitioning

Mining Social-Network Graphs

December 4, 2017 61
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• What is a non-trivial solution to the
equation Lv = 𝜆v for 𝜆=0 (i.e. Lv=0)?

2 3
2 1 1 0 0 0 0
6 1 3 1 1 0 0 07
6 7
6 1 1 2 0 0 0 07
6 7
6 0 1 0 4 1 1 17
6 7
6 0 0 0 1 2 1 07
6 7
4 0 0 0 1 1 3 15
0 0 0 1 0 1 2
Mining Social-Network Graphs

December 4, 2017 62
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Answer
• What is a non-trivial solution to the
equation Lv = 𝜆v for 𝜆=0 (i.e. Lv=0)?

• You already solved it: 1’s (column vector

of constants) 2 3
2 1 1 0 0 0 0
6 1 3 1 1 0 0 07
6 7
6 1 1 2 0 0 0 07
• ACTUALLY, not the 6
6
6 0 1 0 4 1 1
7
17
7
whole story… 6
6
4
0
0
0
0
0
0
1
1
2
1 3
1 077
15
0 0 0 1 0 1 2
Mining Social-Network Graphs

December 4, 2017 63
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• Laplacian of the following graph?

W X

Y Z

Mining Social-Network Graphs

December 4, 2017 64
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Answer
• Laplacian of the following graph?

W X
2 3
1 1 0 0
6 1 1 0 07
6 7
40 0 1 15
Y Z
0 0 1 1
Mining Social-Network Graphs

December 4, 2017 65
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• Null-space of L? (i.e. Lv=0)

W X
2 3
1 1 0 0
6 1 1 0 07
6 7
40 0 1 15
Y Z
0 0 1 1
Mining Social-Network Graphs

December 4, 2017 66
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Answer
• Null-space of L? (i.e. Lv=0)
[1 1 0 0]T
[0 0 1 1]T

W X
2 3
1 1 0 0
6 1 1 0 07
6 7
40 0 1 15
Y Z
0 0 1 1
Mining Social-Network Graphs

December 4, 2017 67
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Connectivity via 0th Eigenvalue

• If nodes i and j are connected, their values in the
corresponding eigenvector must be equal
– This “cancels out” across all rows

• So if the graph is connected, the 0th eigenvector

[c c c … ]T
– Transitivity!

• Otherwise, null-space of L provides connected

components
– Could be detected by a simple clustering algorithm
(e.g. k-Means)

Mining Social-Network Graphs

December 4, 2017 68
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Second-Smallest Eigenvalue
• Let’s call this 𝜆1 (with corresponding v1)

• If 𝜆1 is 0, what do we know?
That there’s not much sense in bi-partitioning :)

• Foreshadowing: from here, we’ll construct

an objective function for a bi-partitioning,
and it will turn out to be minimized
precisely by v1
Mining Social-Network Graphs

December 4, 2017 69
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Setup
• For a given graph, we want to assign each
node to one of two groups
– Let’s represent this assignment, per node, as
the value -1 or +1 for variable ni
– So, a vector, n, of length |V| of either -1 or 1

• And do so in a way that minimizes the

number of graph edges between the
groups - so let’s minimize…
(ni – nj)2 for all edges (i, j)

Mining Social-Network Graphs

December 4, 2017 70
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
What is (ni – nj)2 for two connected nodes that …

• Are in the same group?

(1 - 1)2 = 0; (-1 - -1)2 = 0

• Are in different groups?

(1 - -1)2 = 4; (-1 – 1)2 = 4

So minimizing the sum across edges also

minimizes cross-group edges – woohoo!
Mining Social-Network Graphs

December 4, 2017 71
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Converting to Matrices
• So we have ∑(ni – nj)2
– Or: ∑ (ni2 – 2ninj + nj2)

• The way to express this in terms of the

Laplacian is… nTLn
– Each node will have ni2d(ni) term (diagonal), so
one ni2 per edge = ni2 + nj2
– Each edge will have two (-1)(ninj) terms
(symmetric) = -2ninj
– Each non-existent edge will multiply by zero from
Aij being zero
Mining Social-Network Graphs

December 4, 2017 72
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Checkup
• What is the easiest way to minimize the
following function: ni2 – 2ninj + nj2

Mining Social-Network Graphs

December 4, 2017 73
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Answer
• What is the easiest way to minimize the
following function: ni2 – 2ninj + nj2

• n = [0 0 … 0]
– So we need to force values…

Mining Social-Network Graphs

December 4, 2017 74
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Forcing Non-Trivial Solutions

• Easy: ∑n2 = |V|
– The sum of the assignments must equal the
total number of nodes

• Or, in vector notation nTn=|V|

Mining Social-Network Graphs

December 4, 2017 75
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Objective Function – Minimize!

| |
n Ln + (n n |V |)

Ln n=0

SO, to find a good partitioning we need to...

• Find an eigenvector…
• with the second-smallest eigenvalue
– Why doesn’t 0 work?
Mining Social-Network Graphs

December 4, 2017 76
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (1): Graph

A B

C D

Mining Social-Network Graphs

December 4, 2017 77
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (1): Laplacian

2 3
2 1 1 0
6 1 2 0 17
6 7
4 1 0 1 05
0 1 0 1

Mining Social-Network Graphs

December 4, 2017 78
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (1): Eigendecomposition

Sign = Partitioning (+ vs -)
+ {B, D}
- {A, C}

Mining Social-Network Graphs

December 4, 2017 79
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (1): Partitioning

A B

C D

Mining Social-Network Graphs

December 4, 2017 80
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (2): Graph

Mining Social-Network Graphs

December 4, 2017 81
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (2): Laplacian

Mining Social-Network Graphs

December 4, 2017 82
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (2): Eigendecomposition

Mining Social-Network Graphs

December 4, 2017 83
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Example (2): Partition

Mining Social-Network Graphs

December 4, 2017 84
 CS6220 – Data Mining Techniques･ ･･ Fall 2017･ ･･ Derbinsky

Many Other Topics

• Overlapping communities

• Network/Node properties
– Triangles, neighborhoods
– Centrality, influence

• Network formation

• Problems
– Link prediction
Mining Social-Network Graphs

December 4, 2017 85