label_propagation_communities random_state default change for reproducible partitions

Answer

The current public API of networkx.algorithms.community.label_propagation.label_propagation_communities intentionally leaves random_state=None as the default. That means the visitation order of nodes is chosen by the underlying numpy.random generator each time the function is called, and the resulting community partition can differ between runs on the same graph.

Current API contract

• When random_state is None, the algorithm is nondeterministic.
• Deterministic behaviour is guaranteed only when the caller supplies a reproducible seed (int or numpy.random.RandomState).
• The documentation explicitly states that reproducibility is achieved by providing a seed; it does not promise deterministic output when the seed is omitted.

Why the default should stay nondeterministic

• Changing the default to a fixed seed would alter the public API contract and could break existing code that relies on the stochastic behaviour.
• The internal order of node updates is an implementation detail that may change in future releases; making the output deterministic by default would mask those changes and give a false sense of stability.
• Users who need reproducibility already have a documented, supported path: supply a seed.

How to obtain reproducible partitions

1. Choose a seed value that you can version‑control (e.g., 42 or 2026-10-16).
2. Call the function with that seed:

   import networkx as nx
   from networkx.algorithms.community import label_propagation_communities
   
   G = nx.erdos_renyi_graph(100, 0.05, seed=123)
   communities = list(label_propagation_communities(G, random_state=42))
   

3. Store the seed alongside your graph data or in a configuration file so that others can reproduce the same partition.
4. Optionally, serialize the graph and the seed, then reload and re‑run to confirm consistency across environments.

Documenting reproducibility

• Include the seed value in experiment logs or a dedicated metadata file.
• If you use a numpy.random.RandomState instance, serialize its state (e.g., via np.save or pickle) and restore it before calling the community function.
• State that the algorithm’s output is not guaranteed to remain identical across NetworkX releases even with the same seed; the seed only guarantees determinism for the current implementation.

In short, the default nondeterministic behaviour is part of the API contract. Users who need reproducible partitions should explicitly pass a seed; NetworkX will not change the default to a fixed value in future releases.