You are on page 1of 2

PageRank is an extension of eigen-vector centrality for a directed network.

Let’s first understand the


representation for eigen-vector centrality.

Let’s 𝐶𝑒 (𝑣𝑖 ) be the centrality measure of node, i, represented as a combination of centrality


measures of the nodes having edges incident on node i.
𝑛

𝐶𝑒 (𝑣𝑖 ) = ∑ 𝑎𝑗𝑖 𝐶𝑒 (𝑣𝑗 ) (1)


𝑗=1

To make the summation bounded we normalize the expression of RHS of (1) by a constant 𝜆.
𝑛
1
𝐶𝑒 (𝑣𝑖 ) = ∑ 𝑎𝑗𝑖 𝐶𝑒 (𝑣𝑗 ) (2)
𝜆
𝑗=1

In a directed graph, note that centrality only passes over outgoing edges. Since, in (2) 𝑎𝑗𝑖 = 1 if there
is an edge from node j to i, and 𝑎𝑗𝑖 = 0 otherwise. Therefore, in the presence of directed acyclic
graph centrality becomes zero. Therefore, for the directed graph an extension of eigen-vector
centrality is used known as Katz centrality.
𝑛

𝐶𝑒 (𝑣𝑖 ) = 𝛼 ∑ 𝑎𝑗𝑖 𝐶𝑒 (𝑣𝑗 ) + 𝛽 (3)


𝑗=1

In the directed graph, a node with high centrality value (𝐶𝑒 (𝑣𝑗 )) will pass its high centrality value to
all nodes connected to it. This is undesirable, because everyone having an edge from a well-known
person may not be well-known.

Therefore, we adjust the value of centrality passed by an outgoing link from a node by the number
of out-degree of that node. In other words, we modify (4) in the following way to get PageRank:
𝑛
𝐶𝑒 (𝑣𝑗 )
𝐶𝑒 (𝑣𝑖 ) = 𝛼 ∑ 𝑎𝑗𝑖 +𝛽 (4)
𝑑𝑗𝑜𝑢𝑡
𝑗=1

Note that in matrix notation, using adjacency matrix 𝑨 = {𝑎𝑖𝑗 }, (3) can be expressed as

𝑪𝑒 = 𝛼𝑨𝑇 𝑪𝑒 + 𝛽. 𝟏𝑛 3(𝑎)
𝐶𝑒 (𝑣1 ) 1
Here, 𝑪𝑒 = [ ⋮ ] is a column vector of centrality values, and 𝟏𝑛 = [ ⋮ ]
𝐶𝑒 (𝑣𝑛 ) 𝑛×1 1 𝑛×1

The key to understand the matrix notation for (4) lies in following representation for the weighted
centrality measure.
𝐶𝑒 (𝑣1 ) 1
𝑜𝑢𝑡 𝑜𝑢𝑡 ⋯ 0
𝑑1 𝑑1 𝐶𝑒 (𝑣1 ) 𝑑1𝑜𝑢𝑡 ⋯ 0
−1
⋮ = ⋮ ⋱ ⋮ ×[ ⋮ ] = 𝑫 𝑪𝑒 ; 𝑤ℎ𝑒𝑟𝑒 𝑫 = [ ⋮ ⋱ ⋮ ]
𝐶𝑒 (𝑣𝑛 ) 1 𝐶𝑒 (𝑣𝑛 ) 𝑛×1 𝑜𝑢𝑡
0 ⋯ 𝑑𝑛
0 ⋯ 𝑛×𝑛
[ 𝑑𝑛𝑜𝑢𝑡 ]𝑛×1 [ 𝑑𝑛𝑜𝑢𝑡 ]𝑛×𝑛

Now, the matrix representation of (4) is straightforward using the matrix notation of (3) in (3a)
where 𝑪𝑒 is simply replaced by 𝑫−1 𝑪𝑒 .
𝑪𝑒 = 𝛼𝑨𝑇 𝑫−1 𝑪𝑒 + 𝛽. 𝟏𝑛
With dimensions it can be rewritten as:

𝑪𝑒 (𝑛×1) = 𝛼𝑨𝑇 (𝑛×𝑛) 𝑫−1 (𝑛×𝑛) 𝑪𝑒 (𝑛×1) + 𝛽. 𝟏(𝑛×1)

You might also like