Am 28.07.2018 um 20:54 schrieb Anatol Wegner:
Hi,
I have a question regarding the implementing partial constraints during blockmodel minimization. By partial constraints I mean that I have a group of vertices V1 that are pre-labeled into k clusters i.e. vertices from V1 that are in different clusters should not mix. Then there is a second set of vertices V2 that should be able to mix freely with vertices from both V1 and V2.
I have tried defining a corresponding vertex property and passing it to clabel argument but have always ended up with partitions in which V1 and V2 don't mix in the final state.
You can't use clabel to implement partial constraints. You need to use the 'bfield' that specifies the prior log-probability of a node to belong to a particular group. With this you can set [0, -inf, -inf,...] to pin down a node to group 0, for example. If you use constant values (or leave the vector empty) it will amount to a flat prior, and the placement will be unconstrained. Best, Tiago -- Tiago de Paula Peixoto <tiago@skewed.de>