tau-community-detection 1.3.3

Community detection via Louvain/Leiden + Genetic Algorithm

TAU Community Detection

tau-community-detection implements TAU, an evolutionary community detection algorithm that couples genetic search with Leiden refinements. It is designed for scalable graph clustering with a simple drop-in run_clustering() API, sensible defaults, and multiprocessing support.

---

Highlights

• Evolutionary search: Maintains a population of candidate partitions and applies crossover and mutation tailored for graph clustering.
• Leiden optimization: Refines every candidate with Leiden to ensure modularity gains each generation.
• Multiprocessing aware: Utilises parallel worker pools for population optimization with automatic fallback to sequential mode.
• Fully reproducible: Pass random_seed to seed both TAU's numpy RNG and igraph's Leiden RNG — same seed always produces identical results.
• Input flexibility: Accepts igraph.Graph, networkx.Graph, or a file path. Edge weights are auto-detected.
• Simple API: Use run_clustering(graph) for zero-friction usage, or drop down to TauClustering + TauConfig for full control.

---

Installation

Requires Python 3.10 or newer.

pip install tau-community-detection

To work from a clone:

git clone https://github.com/HillelCharbit/TAU.git
cd TAU
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements-dev.txt
pip install -e .

---

Quick Start

import igraph as ig
from tau_community_detection import run_clustering

g = ig.Graph.Famous("Zachary")

# Zero-friction default usage
clustering = run_clustering(g)
print(f"Communities: {len(clustering)},  Modularity: {clustering.modularity:.4f}")

# Override only the knobs you care about
clustering = run_clustering(
    g,
    resolution_parameter=0.8,
    random_seed=42,
    verbose=True,
    population_size=100,
    max_generations=50,
)

run_clustering() returns an igraph.VertexClustering, so .membership, .modularity, and all standard igraph attributes are available immediately.

NetworkX input

import networkx as nx
from tau_community_detection import run_clustering

g = nx.erdos_renyi_graph(n=500, p=0.02, seed=0)
clustering = run_clustering(g)

Advanced usage with TauClustering

For full control over the lifecycle — including reusing the worker pool across multiple runs:

from tau_community_detection import TauClustering, TauConfig

config = TauConfig(
    resolution_parameter=1.0,
    elite_fraction=0.15,
    immigrant_fraction=0.2,
    stopping_generations=10,
    random_seed=42,
    verbose=True,
)

with TauClustering(g, population_size=60, max_generations=30, config=config) as tau:
    clustering, stats = tau.run(track_stats=True)

print(f"Ran for {len(stats)} generations")
print(f"Final modularity: {clustering.modularity:.4f}")

track_stats=True returns a list of per-generation dicts with keys generation, top_fitness, average_fitness, time_per_generation, convergence, elite_runtime, crossover_runtime.

---

Graph Input

Supported sources:

Type	Notes
igraph.Graph	Passed directly; weights auto-detected from "weight" edge attribute
networkx.Graph	Converted internally; weights auto-detected
str (file path)	Edgelist/NCOL (.graph, .edgelist, .txt) or adjacency list (.adjlist)

For large graphs or high worker counts, passing a file path is recommended — it avoids serialising the graph object across worker processes.

Edge weights are detected automatically. To override:

from tau_community_detection import TauConfig
config = TauConfig(is_weighted=False)   # force unweighted even if file has weights

---

Configuration Reference

All hyperparameters live on TauConfig. Every field is validated on construction — invalid values raise ValueError immediately.

Parameter	Default	Valid range	Description
population_size	60	> 0	Number of candidate partitions per generation
max_generations	500	> 0	Hard cap on evolutionary iterations
worker_count	None	≥ 1	Parallel workers (default: CPU count, capped by population size)
elite_fraction	0.1	(0, 1]	Fraction of best partitions preserved each generation
immigrant_fraction	0.15	(0, 1]	Fraction of fresh random partitions injected each generation
selection_power	5	> 0	Sharpness of fitness-proportional parent selection
elite_similarity_threshold	0.9	[0, 1]	Jaccard threshold below which two elites are considered diverse
stopping_generations	10	> 0	Generations without improvement before early stopping
stopping_jaccard	0.98	[0, 1]	Similarity threshold that counts as "no improvement"
n_iterations	3	> 0	Leiden iterations per fitness evaluation
resolution_parameter	1.0	> 0	Leiden resolution — higher values produce more, smaller communities
sample_fraction_range	(0.2, 0.9)	0 < low ≤ high ≤ 1	Range for random subgraph sampling during population init
is_weighted	None	bool or None	Override weight auto-detection (None = auto)
default_edge_weight	1.0	> 0	Weight assigned to edges when graph is treated as unweighted
weight_attribute	"weight"	str or None	Edge attribute name to read weights from
sim_sample_size	20 000	int or None	Node sample size for Jaccard similarity (None = all nodes)
worker_chunk_size	None	int or None	Tasks per worker per batch (None = auto)
reuse_worker_pool	True	bool	Keep worker pool alive between calls on the same instance
random_seed	None	int or None	Seeds both numpy and igraph's Leiden RNG for fully deterministic results
verbose	False	bool	Log progress to the standard Python logger

run_clustering() exposes the most common parameters directly. Any TauConfig field can also be passed as a keyword argument:

clustering = run_clustering(g, elite_fraction=0.2, stopping_generations=5)

---

Development

pip install -r requirements-dev.txt
pip install -e .
make lint     # ruff checks
make test     # pytest
make coverage # pytest + coverage report
make build    # build sdist + wheel

Continuous Integration

GitHub Actions runs lint, tests (Python 3.10 and 3.11), and a package build on every push and pull request. Set the CODECOV_TOKEN secret to upload coverage reports.

Publishing

1. Bump version in setup.cfg and commit.
2. Tag the release: git tag vX.Y.Z && git push --tags.
3. Run the Publish Package workflow. Use TEST_PYPI_API_TOKEN for a dry run on TestPyPI, or PYPI_API_TOKEN to publish to PyPI.

---

Reference & Citation

If you use TAU in your research, please cite:

From Leiden to Tel-Aviv University (TAU): exploring clustering solutions via a genetic algorithm Gal Gilad and Roded Sharan. PNAS Nexus, Volume 2, Issue 6, June 2023. DOI: 10.1093/pnasnexus/pgad180

@article{gilad2023tau,
  title={From Leiden to Tel-Aviv University (TAU): exploring clustering solutions via a genetic algorithm},
  author={Gilad, Gal and Sharan, Roded},
  journal={PNAS Nexus},
  volume={2},
  number={6},
  pages={pgad180},
  year={2023},
  publisher={Oxford University Press}
}

---

License

MIT License © 2023 Hillel Charbit