Main paper can be found here.

recursive and procedural approach to building real-world-like graphs
Erdos-Renyi model becomes a special case of this approach (with a = b = c = d = 0.25)
proposes an AutoMAT-fast parameter fitting algorithm
generates graphs with:
- power-law distribution
- directed, undirected or bipartite properties
- very efficient $$O(E log(E) log(N))$$ complexity
- can also generate "winner does not take all" kind of graphs too
- generate communities (and also sub-communities within communities)
- smaller diameter (proved experimentally in this paper)

Parameters:
- $$N$$ - number of nodes in the graph
- $$E$$ - number of edges in the graph
- $$2^n$$ - number of nodes in the RMAT graph (where $$n = ceil(log2(N))$$)
- $$a, b, c, d$$ - probabilities of each quadrant in the adjacency matrix
- $$a + b + c + d = 1$$
Typically: $$a \ge b$$ and $$a \ge c$$ and $$a \ge d$$
Main algo
- pick the quadrant based on the probabilities $$a, b, c, d$$
- keep doing this until we hit the "leaf node" which is the pair of nodes which will get an edge
- repeat this until we have $$E$$ edges
to smooth out fluctuations in the degree distributions, these probabilities are perturbed (and renormalized) slightly at each level