Generate synthetic networks¶
It is necessary to benchmark our algorithm via synthetic networks, where we have perfect confidence
on the model parameters. Here we can control the nodes’ block membership \(b\), the edge count matrix \(e_{rs}\),
and the degree distribution \(d\). Once we specified these 3 parameters,
we call Graph-tool’s
generate_sbm function to
generate a graph-tool graph instance,
and then use that instance as an input for our biSBM.optimalks.OptimalKs class.
Block membership \(b\)¶
Edge count matrix \(e_{rs}\)¶
Easy cases¶
The easy cases are bipartite version of the planted partition models. It is a \((K_a, K_b)\)-bipartite structure with \(e_{rs} = p c / K + (1 - p)(1 - c)/K(K-1)\), where \(K_a = K_b = K\) and \(p = c_{out}/c_{in}\) is a parameter that controls how crisp the structure is. When \(p = 0\), we expect no connections at all between nodes of different groups, i.e., \(e_{rs} = E \delta_{rs}/ K\).
.png)
A easy case with \(\langle k\rangle = 10\) and \(p=0.1\), planted as \((K_a, K_b) = (5, 5)\) equally-sized communities.¶
Hard cases¶
The hard cases are planted networks where \(K_a \neq K_b\), but the edge count matrix is still designed in a way that allow the control of network structural strength.
.png)
A hard case with \(\langle k\rangle = 10\) and \(p=0.1\), planted as \((K_a, K_b) = (5, 31)\) equally-sized communities.¶
Even harder cases¶
A even harder case is simply a random draw out of the ensemble of edge count matrices where we only fix \(K_a\) and \(K_b\).
.png)
A harder case with \(\langle k\rangle = 10\), planted as \((K_a, K_b) = (5, 31)\) equally-sized communities. This is the strongest structure out of \(10^6\) random \(e_{rs}\) draws.¶