GraRep Algorithm | |
Input | |
Adjacency matrix S on graph | |
Maximum transition step K | |
Log shifted factor β | |
Dimension of representation vector d | |
1. Get k-step transition probability matrix Ak | |
Compute A = D−1S | |
Calculate A−1, A−2,  A−3, …,  Ak, respectively | |
2. Get each k-step representations | |
For k = 1 to K | |
 2.1 Get positive log probability matrix | |
 calculate \( {\Gamma}_1^k,{\Gamma}_2^k,{\Gamma}_3^k,\dots, {\Gamma}_N^k \) (\( {\Gamma}_i^k={\sum}_p{A}_{p,j}^k \)) respectively | |
 calculate \( \left\{{X}_{i,j}^k\right\} \) | |
  \( {X}_{i,j}^k \) = log (\( \frac{A_{i,j}^k}{\Gamma_j^k} \)) – log(β) | |
 assign negative entries of Xk to 0 | |
 2.2 Construct the representation vector Wk | |
 SVD(Xk) = [Uk, ∑k, (Vk)T] | |
 \( {W}^k={U}_d^k\ {\left({\sum}_d^k\right)}^{\frac{1}{2}} \) | |
End for | |
3. Concatenate all the k-step representations | |
W = [W1, W2, W3, …Wk] | |
Output | |
Matrix of the graph representation W |