Community Detection
Community detection groups nodes under a structural objective. It does not discover a unique natural truth. Modularity compares within-group edge weight with a degree-preserving null expectation, and algorithms such as Louvain heuristically search a large space of partitions.
Prerequisites
Complete centrality and its failure modes. Be prepared to justify every conversion from the directed weighted multigraph.
Learning goals
- Describe the graph representation and null model behind modularity.
- Show how the resolution parameter changes a partition.
- Recognize the resolution limit, heuristic randomness, and near-optimal degeneracy.
- Treat community labels as exploratory summaries rather than identities or causes.
Analysis contract
- Graph type: an undirected weighted simple projection; reciprocal and parallel event strengths are summed, self-loops are removed because they are events rather than pre-aggregated communities, and Gia remains an isolate.
- Normalization: weighted modularity compares observed within-community strength with the configuration null model at declared resolution ; no cross-resolution score comparison is treated as absolute.
- Disconnected case: each isolate appears as a singleton, and disconnected components cannot be joined by observed edges; component-wise analysis may be clearer in larger data.
- Interpretation limit: a partition and its score are evidence, not ground-truth groups, identity, cohesion, or causality.
Resolution and degeneracy
For modularity, lower favors larger groups and higher favors smaller groups. The resolution limit means modularity at a fixed scale can merge small, otherwise intuitive communities. Conversely, high resolution can fragment useful structure.
Louvain is a heuristic. Node order and its seeded random shuffle can alter the returned local optimum. Many distinct partitions can have equal or nearly equal objective values—degeneracy—so a single run should not be presented as unique. Sensitivity across seeds, resolution values, graph construction, and plausible missing edges matters more than a colorful partition plot.
Worked example
import csv
from pathlib import Path
import networkx as nx
fixture = Path("tests/fixtures/graph-data-science")
with (fixture / "interaction-nodes.csv").open(encoding="utf-8") as handle:
node_ids = [row["node_id"] for row in csv.DictReader(handle)]
with (fixture / "interaction-relationships.csv").open(encoding="utf-8") as handle:
rows = [row for row in csv.DictReader(handle) if row["record_status"] == "valid"]
graph = nx.Graph()
graph.add_nodes_from(node_ids)
for row in rows:
source, target = row["source"], row["target"]
if source == target:
continue
strength = float(row["weight"])
if graph.has_edge(source, target):
graph[source][target]["strength"] += strength
else:
graph.add_edge(source, target, strength=strength)
def ordered(partition):
return sorted((sorted(group) for group in partition), key=lambda group: (group[0], len(group)))
for resolution in (0.1, 1.0, 3.0):
partition = nx.community.louvain_communities(
graph, weight="strength", resolution=resolution, seed=53
)
print(f"resolution_{resolution:.1f}={ordered(partition)}")
Expected output:
resolution_0.1=[['ada', 'ben', 'cy', 'dia', 'eli', 'fox'], ['gia']]
resolution_1.0=[['ada', 'ben', 'fox'], ['cy', 'dia', 'eli'], ['gia']]
resolution_3.0=[['ada'], ['ben'], ['cy'], ['dia', 'eli'], ['fox'], ['gia']]
These are three answers to three objectives, not three estimates converging on one hidden label set.
Limitations
- The undirected projection discards direction; summing reciprocal events changes the null model.
- Louvain can return poorly connected communities, and fixed seed gives repeatability rather than truth.
- The resolution limit and partition degeneracy undermine claims of a uniquely correct scale.
- Missing or erroneous edges can move nodes or merge groups; the fixture's missing relationship stays unknown.
Practice
Run several seeds at and compare partitions and modularity values. Then add one explicitly hypothetical bridge and describe the sensitivity without calling either result correct.
Next steps
Continue to bipartite and multipartite networks before projecting data that naturally contains more than one node type.
Primary sources
Checked 2026-10-18: NetworkX, louvain_communities; NetworkX, modularity; V. D. Blondel et al., Fast unfolding of communities in large networks; S. Fortunato and M. Barthélemy, Resolution limit in community detection.