pyCombo is a python wrapper around C++ implementation of the [network] community detection algorithm called "Combo".
Details of the algorithm are described in the paper "General optimization technique for high-quality community detection":
Sobolevsky, S., Campari, R., Belyi, A. and Ratti, C., 2014. General optimization technique for high-quality community detection in complex networks. Physical Review E, 90(1), p.012811.
You can install the latest release of pycombo directly from PyPI:
python -m pip install pycomboPre-built wheels are published for Linux (x86_64, aarch64), macOS (Intel + Apple Silicon), and Windows.
Starting with v1.2, only Python 3.9+ is supported. For older Python versions, install the last compatible release:
python -m pip install pycombo==0.1.08 # Python 3.8
python -m pip install pycombo==0.1.07 # Python 3.7Partition a NetworkX graph and get the modularity score:
import networkx as nx
import pycombo
G = nx.karate_club_graph()
partition, modularity = pycombo.execute(G, random_seed=42)
print(f"Found {len(set(partition.values()))} communities, modularity={modularity:.4f}")Write community labels back onto the graph nodes:
partition, modularity = pycombo.execute(
G,
random_seed=42,
community_attribute="community",
)
assert G.nodes[0]["community"] == partition[0]Return a cdlib clustering for comparison with other methods:
from cdlib import algorithms
combo_clustering, modularity = pycombo.execute(G, random_seed=42, as_clustering=True)
leiden_clustering = algorithms.leiden(G)Package supports NetworkX graphs, Pajek .net files, and adjacency matrices passed as numpy array or list.
Combo algorithm uses modularity score as a loss function, but you can use your own metrics as edge weights with treat_as_modularity=True parameter.
- graph :
nx.Graphobject, or string treated as path to Pajek.netfile. - weight :
Optional[str], defaults toweight. Graph edges property to use as weights. IfNone, graph assumed to be unweighted. Ignored if graph is passed as string (path to the file), or such property does not exist. - max_communities :
Optional[int], defaults toNone. Maximum number of communities. If <= 0 or None, assume to be infinite. - modularity_resolution :
float, defaults to 1.0. Modularity resolution parameter. - num_split_attempts :
int, defaults to 0. Number of split attempts. If 0, autoadjust this number automatically. - fixed_split_step :
int, defaults to 0. Step number to apply predefined split. If 0, use only random splits. if >0, sets up the usage of 6 fixed type splits on every fixed_split_step. - start_separate : bool, default False. Indicates if Combo should start from assigning each node into its own separate community. This could help to achieve higher modularity, but it makes execution much slower.
- treat_as_modularity : bool, default False. Indicates if edge weights should be treated as modularity scores. If True, the algorithm solves clique partitioning problem over the given graph, treated as modularity graph (matrix). For example, this allows users to provide their own custom 'modularity' matrix.
modularity_resolutionis ignored in this case. - verbose : int, defaults to 0. Indicates how much progress information Combo should print out. For now Combo has only one level starting at verbose >= 1.
- intermediate_results_path : Optional str, defaults to None. Path to the file where community assignments will be saved on each iteration. If None or empty, intermediate results will not be saved.
- return_modularity : bool, defaults to
True. Indicates if function should return achieved modularity score. - random_seed : int, defaults to None. Random seed to use. None indicates using some internal default value that is based on time and is expected to be different for each call.
- community_attribute : Optional str. When partitioning a NetworkX graph, write labels to
graph.nodes[node][community_attribute]. - as_clustering : bool, defaults to
False. Return acdlib.classes.NodeClusteringinstead of a dict (requires cdlib).
- partition :
Dict{int : int}, community labels for each node. - modularity :
float. Achieved modularity value. Only returned ifreturn_modularity=True.
More examples can be found in example folder.
This repo uses C++ source as a git submodule.
So for local development, clone with --recurse-submodules flag, as:
git clone --recurse-submodules https://github.com/Casyfill/pyComboOr, if you've already cloned it without --recurse-submodules, run:
git submodule update --init --recursivePackage is built and managed via uv.
- To use a specific Python version run
uv python pin 3.13. - To install dev dependencies, run
uv sync. - To build distributions run
uv build. - To build all platform wheels locally run
uv run cibuildwheel --output-dir wheelhouse. - To run tests execute
uv run pytest.
pyCombo is licensed under the GNU General Public License v3.0 or later (GPLv3+).