GML Floating-Point Attribute Coercion in NetworkX readwrite.gml

Answer

Confirmed facts from NetworkX’s GML reader:

• The parser attempts conversions in the order int → float → bool → string and stores the first successful type.
• A floating‑point value is coerced to an integer when its GML token lacks a decimal point or exponent (e.g., the value 3.0 written as 3). The fractional part is lost.
• Both nx.Graph and nx.MultiGraph use the same read_gml implementation, so their behavior is identical for node attributes.
• When several node attributes violate the GML type system, each attribute is processed independently; there is no cross‑attribute prioritization.

Likely explanation

The coercion occurs because the GML writer may drop the fractional part when a float happens to be an exact integer (or when custom formatting strips the decimal). The reader then sees a token like 3 and successfully converts it with int() before trying float(), so the original 3.0 becomes an integer 3.

Verification steps

1. Create a graph and add a node attribute with a float that is an integer value:

   import networkx as nx
   G = nx.Graph()
   G.add_node(0, weight=3.0)   # float
   nx.write_gml(G, 'test.gml')
   H = nx.read_gml('test.gml')
   print(type(H.nodes[0]['weight']))  # ->  if token lacks decimal
   

2. Repeat with nx.MultiGraph and confirm the same result.
3. Add two attributes, one written as 2.5 (keeps decimal) and another as 4 (no decimal), read back and verify the first stays float while the second becomes int.

If you observe a different outcome, please confirm your NetworkX version (e.g., nx.__version__) as older releases may have divergent parsing logic.