Graph Theory: Blockmodeling

Role Structure and Block Relations

Part 1: The Correlation Distance & CONCOR

The Correlation Distance (\(d^{corr}_{k,l}\))

  • Beyond Euclidean distance, we can measure structural similarity using the Pearson Correlation Coefficient (\(r_{kl}\) or \(d^{corr}_{kl}\)) between node vectors.
  • Compares row (or column) vectors and returns a value between \(-1.0\) and \(+1.0\):
    • \(+1.0\): Nodes are perfectly structurally equivalent.
    • Positive (\(> 0.0\)): Nodes are structurally similar.
    • Negative (\(< 0.0\)): Nodes are structurally dissimilar.
    • \(-1.0\): Nodes are completely different, pointing to completely disjoint sets of neighbors.
  • The formula is:
    • \[d^{corr}_{k, l} = \frac{\sum_j (a_{(k)j} - \overline{a}_{(k)j}) \times (a_{(l)j} - \overline{a}_{(l)j})}{\sqrt{\sum_j (a_{(k)j} - \overline{a}_{(k)j})^2 \times \sum_j (a_{(l)j} - \overline{a}_{(l)j})^2}}\]

Meaning of Row Means (\(\overline{a}_{(k)j}\))

  • What does \(\overline{a}_{(k)j}\) represent?
    • It is simply the average of the row elements for node \(k\).
    • \[\overline{a}_{(k)j} = \frac{\sum_j a_{kj}}{N}\]
    • Where \(N\) is the number of columns (the order of the graph).
  • For a binary adjacency matrix, the row mean is simply the node degree divided by the number of nodes:
    • Node \(A\) has ties to \(\{C, D\}\) in a 6-node network:
    • Vector \(a_{(A)j} = (0, 0, 1, 1, 0, 0)\).
    • Average: \(\overline{a}_{(A)j} = \frac{2}{6} \approx 0.33\).

Mean Differences Vector

  • Subtracting the row mean \(\overline{a}_{(A)j} \approx 0.33\) from each entry of \(A\)’s row vector \(a_{(A)j} = (0, 0, 1, 1, 0, 0)\) gives the Mean Differences vector:
A B C D E F
Row A 0 0 1 1 0 0
A - Mean (0 - 0.33) (0 - 0.33) (1 - 0.33) (1 - 0.33) (0 - 0.33) (0 - 0.33)
Difference -0.33 -0.33 0.67 0.67 -0.33 -0.33
  • We compute these difference vectors for all pairs of nodes, multiply them in the numerator, normalize them in the denominator, and obtain the correlation matrix.

Correlation Distance Matrix (\(S^{corr}\))

A B C D E F
A -- 1 -0.71 -0.71 0.71 -0.32
B 1 -- -0.71 -0.71 0.71 -0.32
C -0.71 -0.71 -- 1 -1 0.45
D -0.71 -0.71 1 -- -1 0.45
E 0.71 0.71 -1 -1 -- -0.45
F -0.32 -0.32 0.45 0.45 -0.45 --
  • Structurally equivalent pairs (\(\{A, B\}\) and \(\{C, D\}\)) have a correlation of 1.00.
  • Completely disjoint pairs (\(\{C, E\}\) and \(\{D, E\}\)) are perfectly negatively correlated (-1.00).
  • Similar but non-equivalent nodes (\(C\) and \(F\)) have a correlation of 0.45.

Iterated Correlation Distances: CONCOR

  • What happens if we take the correlation distance of a correlation distance matrix?
  • CONCOR (Convergence of Iterated Correlations; Breiger, Boorman, & Arabie, 1975):
    • If we recursively calculate the correlation distance matrix of the previous correlation matrix, the values mathematically converge to only \(+1.0\) and \(-1.0\).
    • This slices the network into two distinct, highly clear-cut blocks of structurally equivalent or similar nodes.
  • Let’s look at this mathematical convergence in action.

Iterative Convergence (2nd & 3rd Iterations)

2nd Iteration Matrix

  • Values begin polarizing toward \(+1\) and \(-1\).
A B C D E F
A -- 1 -0.97 -0.97 0.97 -0.84
B 1 -- -0.97 -0.97 0.97 -0.84
C -0.97 -0.97 -- 1 -1 0.84
D -0.97 -0.97 1 -- -1 0.84
E 0.97 0.97 -1 -1 -- -0.84
F -0.84 -0.84 0.84 0.84 -0.84 --

3rd Iteration Matrix

  • Every cell reaches exactly \(+1.00\) or \(-1.00\)!
A B C D E F
A -- 1 -1 -1 1 -0.99
B 1 -- -1 -1 1 -0.99
C -1 -1 -- 1 -1 0.99
D -1 -1 1 -- -1 0.99
E 1 1 -1 -1 -- -0.99
F -0.99 -0.99 0.99 0.99 -0.99 --

Structural Positions via CONCOR

A B C D E F
A 1.00 1.00 -1.00 -1.00 1.00 -0.99
B 1.00 1.00 -1.00 -1.00 1.00 -0.99
C -1.00 -1.00 1.00 1.00 -1.00 0.99
D -1.00 -1.00 1.00 1.00 -1.00 0.99
E 1.00 1.00 -1.00 -1.00 1.00 -0.99
F -0.99 -0.99 0.99 0.99 -0.99 1.00
  • By analyzing the rows of \(+1\) and \(-1\), we extract perfect structurally equivalent positions:
    • Position 1 (Salmon): Nodes \(\{A, B\}\)
    • Position 2 (Green): Nodes \(\{C, D\}\)
    • Position 3 (Blue): Node \(\{E\}\)
    • Position 4 (Purple): Node \(\{F\}\)

Part 2: Blockmodeling a 22-Node Network

Uncovering Secret Patterns in Large Networks

  • Let’s apply CONCOR to a larger, messier 22-node network on the right.
  • It contains:
    • Three high-degree hub nodes (\(A\), \(B\), \(C\)).
    • Sets of identical peripheral leaf nodes connected to each hub.
    • A highly central node \(D\) bridging hubs.
    • Two peripheral nodes \(U, V\) connected only to \(S\).
  • Let’s see how CONCOR recursively partitions this graph.

Step 1: First Split (Two Giant Blocks)

  • Running CONCOR on the whole 22-node network splits it into two massive blocks:
    • Block 1: \(\{V, U, S, H, G, F, E, T, C, A, B\}\) (Colored Red)
    • Block 2: \(\{R, Q, P, O, N, M, L, K, J, D, I\}\) (Colored Blue)

Step 2: Going Deeper (The Hierarchical Partition)

  • Running CONCOR recursively on the induced subgraphs of each block yields nested subdivisions:
    • Block 1 Splits:
      • Block 1a (Core hub/bridge): \(\{A, B, C, S, U, V\}\)
      • Block 1b (A’s peripheral leaves): \(\{E, F, G, H, T\}\)
    • Block 2 Splits:
      • Block 2a (B’s peripheral leaves): \(\{I, J, K, L, M\}\)
      • Block 2b (C’s peripheral leaves): \(\{N, O, P, Q, R\}\)
      • Block 2c (Unique central bridge): \(\{D\}\) (Occupies a block entirely by itself!)
  • Further splits on Block 1a isolate \(\{U, V\}\), and divide the remaining \(\{A, B, C, S\}\) into \(\{A, S\}\) and \(\{B, C\}\).

Final Seven-Block Partition

  • The final, elegant block partition reveals how role structures are naturally isolated:

The Blocked Adjacency Matrix (Blocks 1–3)

I J K L M N O P Q R T E F G H B C A S U V D
I 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
J 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
K 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
L 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
M 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
N 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
O 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
P 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
Q 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
R 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
T 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0
E 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
F 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
G 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
H 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

The Blocked Adjacency Matrix (Blocks 4–7)

I J K L M N O P Q R T E F G H B C A S U V D
B 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
C 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 1
A 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 1
S 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 1 0
U 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
V 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
D 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0
  • Thick black lines divide the matrix into block submatrices.
  • Diagonal squares represent within-block ties; off-diagonal rectangles represent between-block ties.
  • We identify Zero Blocks (filled only with 0s) and One Blocks (containing some 1s).

Part 3: Image Matrices & Reduced Graphs

Reducing Complexity to an Image Matrix

  • The blocked adjacency matrix can be simplified to a reduced image matrix (\(\mathbf{B}\)) of dimensions \(b \times b\) (where \(b = 7\) is the number of blocks).
  • Rule of Reduction:
    • We record a 0 if the relation block is a Zero Block (completely disconnected).
    • We record a 1 if the relation block is a One Block (contains ties).
  • This yields a highly simplified map of role interactions:

\[ \mathbf{B} = \begin{pmatrix} \text{Zero} & \text{Zero} & \text{Zero} & \text{One} & \text{Zero} & \text{Zero} & \text{Zero} \\ \text{Zero} & \text{Zero} & \text{Zero} & \text{One} & \text{Zero} & \text{Zero} & \text{Zero} \\ \text{Zero} & \text{Zero} & \text{Zero} & \text{Zero} & \text{One} & \text{Zero} & \text{Zero} \\ \text{One} & \text{One} & \text{Zero} & \text{Zero} & \text{Zero} & \text{Zero} & \text{One} \\ \text{Zero} & \text{Zero} & \text{One} & \text{Zero} & \text{Zero} & \text{One} & \text{One} \\ \text{Zero} & \text{Zero} & \text{Zero} & \text{Zero} & \text{One} & \text{Zero} & \text{Zero} \\ \text{Zero} & \text{Zero} & \text{Zero} & \text{One} & \text{One} & \text{Zero} & \text{Zero} \end{pmatrix} \]

Image Matrix Table

Leaves_I Leaves_N Leaves_T Hubs_BC Hubs_AS Leaves_UV Bridge_D
Leaves_I 0 0 0 1 0 0 0
Leaves_N 0 0 0 1 0 0 0
Leaves_T 0 0 0 0 1 0 0
Hubs_BC 1 1 0 0 0 0 1
Hubs_AS 0 0 1 0 0 1 1
Leaves_UV 0 0 0 0 1 0 0
Bridge_D 0 0 0 1 1 0 0
  • High-level representation of block relations.
  • Notice that Leaf blocks do not connect to themselves (0s on diagonal), but connect specifically to their respective Hub blocks.

Reduced Graph Representation

  • We can plot the image matrix as a reduced graph.
  • The Power of Blockmodeling:
    • Distills a sprawling, complex network of 22 nodes down to a highly intuitive role-interaction model of just 7 block-nodes.
    • Revealing the structural blueprint of the system: peripheral elements (leaves) feeding into coordinated brokers (hubs), bridged by a highly central node (\(D\)).

Blockmodeling: A Structural Summary

  • Blockmodeling Distills Structure:
    • Slices complex social structures down to high-level system blueprints.
  • CONCOR algorithm:
    • Recursively polarizes Pearson correlations to group similar actors.
  • Zero vs. One Blocks:
    • Blocked adjacency matrices cleanly isolate patterns of connection (One blocks) and absolute separation (Zero blocks).
  • Reduced Graph Models:
    • Image matrices represent relations between roles, creating simplified reduced graphs that make networks easy to read, interpret, and model.