Choice of hierarchy level augmentation in cookbook
Hi Tiago, In your cookbook examples on hierarchical SBMs I have noticed that frequently you augment the hierarchy to ten levels. Is there a particular reason for that choice? Best wishes, Philipp -- Sent from: http://main-discussion-list-for-the-graph-tool-project.982480.n3.nabble.com/
On 24.04.2018 12:32, P-M wrote:
In your cookbook examples on hierarchical SBMs I have noticed that frequently you augment the hierarchy to ten levels. Is there a particular reason for that choice?
The number of levels should be large enough to match the inferred depth, i.e. the last levels should have only one group most of the time. A value of L=10 is good enough for most purposes. Best. Tiago -- Tiago de Paula Peixoto <tiago@skewed.de>
Thank you. I take it that having several single group levels at the "top" is fine and that the algorithm does not automatically increase the number of levels if we just stick with the state found by using gt.minimize_nested_blockmodel_dl? What are the consequences of having insufficient levels and what are the consequences of having excessive levels? Is it possible to make general judgements about that? -- Sent from: http://main-discussion-list-for-the-graph-tool-project.982480.n3.nabble.com/
On 24.04.2018 14:11, P-M wrote:
Thank you. I take it that having several single group levels at the "top" is fine and that the algorithm does not automatically increase the number of levels if we just stick with the state found by using gt.minimize_nested_blockmodel_dl? What are the consequences of having insufficient levels and what are the consequences of having excessive levels? Is it possible to make general judgements about that?
It doesn't automatically adapt the maximum size of the hierarchy. If the maximum size is too small, then the model might underfit (albeit probably only very slightly). It can't be too large, so it's better to be safe and give it enough room. -- Tiago de Paula Peixoto <tiago@skewed.de>
OK, thank you for the clarification. Best, Philipp -- Sent from: http://main-discussion-list-for-the-graph-tool-project.982480.n3.nabble.com/
participants (2)
-
P-M -
Tiago de Paula Peixoto