Unexpected values in NetworkX to_numpy_array output for parallel edges

Direct Answers

1. Aggregation rule: to_numpy_array sums the weights of parallel edges by default when a weight attribute is provided for MultiGraph or MultiDiGraph.
2. Documentation and configurability: Since NetworkX 2.5 (2019), the multigraph_weight parameter accepts a callable (sum, min, max, or a custom function) to control aggregation. The default is sum. This parameter appears in the function signature and docstring but is easy to miss in narrative tutorials.
3. Downstream analysis assumption: Do not assume the default sum behavior without verification. Explicitly set multigraph_weight to make intent clear and reproducible.

Confirmed Behavior

For a MultiGraph with two parallel edges between nodes 0 and 1 carrying weights 2 and 5:

import networkx as nx
import numpy as np

G = nx.MultiGraph()
G.add_edge(0, 1, weight=2)
G.add_edge(0, 1, weight=5)

# Default (sum)
A = nx.to_numpy_array(G, weight='weight')
print(A[0, 1])  # 7.0

# Explicit max
A_max = nx.to_numpy_array(G, weight='weight', multigraph_weight=max)
print(A_max[0, 1])  # 5.0

# Explicit min
A_min = nx.to_numpy_array(G, weight='weight', multigraph_weight=min)
print(A_min[0, 1])  # 2.0

The same logic applies to MultiDiGraph (directed parallel edges are aggregated per ordered node pair). For Graph/DiGraph (no parallel edges), weight simply retrieves the single edge's attribute value.

Verification Steps

1. Create a minimal MultiGraph with two parallel edges of known distinct weights.
2. Call to_numpy_array(G, weight='weight') and confirm the matrix entry equals the sum.
3. Repeat with multigraph_weight=max and multigraph_weight=min to verify the parameter works.
4. Check your NetworkX version: nx.__version__ (requires ≥2.5 for multigraph_weight).
5. Consult the official API docs: to_numpy_array reference.

Cautions

• Default sum can silently produce values that match no individual edge weight, misleading pathfinding, centrality, or other numeric analyses.
• Versions <2.5 lack multigraph_weight and always sum; code relying on the default may behave differently across environments.
• Custom aggregation functions must accept an iterable of weights and return a single scalar.
• Self-loops are included on the diagonal and aggregated the same way.
• If the weight attribute is missing on some edges, NetworkX 3.0+ uses a default of 1 unless overridden via the weight parameter's default argument.

Recommendation

Always specify multigraph_weight explicitly in production code (e.g., multigraph_weight=sum for the documented default). Add a unit test that validates the aggregation against a small fixture graph to guard against version changes.