Draft PR: Add modularity and modularity_adata functions to scanpy.metrics #3613
Conversation
There was a problem hiding this comment.
Hi! Apart from the issue with igraph being an optional dependency, this looks good!
I think we have get_igraph_from_adjacency which might be useful, but maybe not.
Could you please add tests? We have many examples on how, best would probably be
- for the direct variant, manually create very small graphs to run this on so you can be sure the results are correct
- for the anndata version, use
neighborsto create the connectivity matrix.
Please add @needs.igraph so the test only runs when igraph is installed
if you’re unsure about anything, please search the code for examples or ask me!
If you end up implementing a non-igraph flavor for this, please test using parametrization, e.g.: @pytest.mark.parametrize("directed", [True, False], ids=["directed", "undirected"])
| @@ -89,3 +95,61 @@ def confusion_matrix( | |||
| df = df.loc[np.array(orig_idx), np.array(new_idx)] | |||
|
|
|||
| return df | |||
|
|
|||
|
|
|||
| def modularity(connectivities, labels, mode="UNDIRECTED") -> float: | |||
There was a problem hiding this comment.
This comment also applies to the other function.
please use type annotations and remove the : free-text type from the parameters in the docstring.
E.g. instead of connectivities : array-like or sparse matrix in the docstring, it should be connectivities: ArrayLike | CSBase in the function definition. (search the code for examples if you’re unsure where to import them from).
Also since there are limited options, use Literal["UNDIRECTED", "DIRECTED"] instead of str. Text for defaults is added automatically, so don‘t write them into the docstring either. But here a is_directed: bool parameter would be better anyway, there will never be more than two options.
| @@ -4,11 +4,15 @@ | |||
|
|
|||
| from typing import TYPE_CHECKING | |||
|
|
|||
| import igraph as ig | |||
There was a problem hiding this comment.
This is an optional dependency and therefore can’t be imported on the top-level.
Optimally we’d have multiple implementations so this works without igraph, but it’s OK to just import igraph in the function and have it fail when igraph can’t be imported.
| from natsort import natsorted | ||
| from pandas.api.types import CategoricalDtype | ||
|
|
||
| from scanpy._compat import CSRBase |
There was a problem hiding this comment.
please use relative imports (from .._compat)
|
|
||
|
|
||
| def modularity(connectivities, labels, mode="UNDIRECTED") -> float: | ||
| # default mode is undirected?? can be specified as directed or undirected |
| if isinstance(connectivities, CSRBase): | ||
| # Convert sparse matrix to dense format so that igraph can handle it | ||
| # Weighted_Adjacency expects with nested lists or numpy arrays and not sparse matrices | ||
| dense_connectivities = connectivities.toarray() |
There was a problem hiding this comment.
The graphs we use are inherently sparse, so we should use APIs that operate on sparse graphs and never densify if possible.
Are you sure there’s no support for initializing igraph from a sparse matrix?
| if isinstance(labels, pd.Series): | ||
| labels = labels.values | ||
| # making sure labels are in the right format, i.e., a list of integers | ||
| labels = pd.Categorical(np.asarray(labels)).codes |
There was a problem hiding this comment.
that’s a lot of converting around! Are you sure you can’t just pass labels into pd.Categorical directly?
| label_array = adata.obs[labels] if isinstance(labels, str) else labels | ||
| # extracting the connectivities from adata.obsp["connectivities"] | ||
| connectivities = adata.obsp[obsp] | ||
| return modularity(connectivities, label_array) |
There was a problem hiding this comment.
Hmm, we don’t currently store if the connectivity calculation assumes a directed or undirected graph …
Just a thought.
This adds two functions to compute modularity scores from a given graph and a clustering like Leiden or Louvain. The goal is to make it easier to compare different community detection methods using an external metric. This follows up on issue #2908. To my knowledge, there is no built-in way to compare clustering results nor ways to calculate modularity score.
Functions: