pyCOMBO

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.

Installation

You can install the latest release of pycombo directly from PyPI:

python -m pip install pycombo

Pre-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 3.7

Quick Start

Partition 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.

Parameters

• graph : nx.Graph object, or string treated as path to Pajek .net file.
• weight : Optional[str], defaults to weight. Graph edges property to use as weights. If None, 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 to None. 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_resolution is 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 a cdlib.classes.NodeClustering instead of a dict (requires cdlib).

Returns

• partition : Dict{int : int}, community labels for each node.
• modularity : float. Achieved modularity value. Only returned if return_modularity=True.

More examples can be found in example folder.

Development

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/pyCombo

Or, if you've already cloned it without --recurse-submodules, run:

git submodule update --init --recursive

Package 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.

License

pyCombo is licensed under the GNU General Public License v3.0 or later (GPLv3+).

Information

• project web_site
• PyCombo paper (Arxiv)