The calculated metric for the quality of the overlapping community segmentation results is incorrect

[destiny1009]

Using the simple community division case to calculate the overlapping NMI value gives a result of 0, but in fact, the result is not 0.

Second, when using the f1-score metric, I found that I could not get a result and simply quit the program. Is it because I made a mistake in using it?

Third, regarding the overlapping communities, only the graph. cover type can be used to calculate the overlapping related metrics. But the graph.cover type cannot be obtained by reading the file, only manual code can be written to read the file and then add them one by one, this part can be optimized perhaps?

[destiny1009]

Can anyone help me explain this?

[fabratu]

Hi,

and sorry for the delayed response. Concerning your first question. Manually setting the cover includes some not really intuitive procedures (we should improve the documentation in this regard). The easiest way is to create a singleton clustering and remove nodes (if that's applicable for your example) from their singleton subsets.

Sample code:

import networkit as nk

G = nk.Graph(5)
G.addEdge(0,1)
G.addEdge(1,2)
G.addEdge(2,3)
G.addEdge(3,4)

f1= nk.Cover(n=5)
f1.allToSingletons()
print(f1.getSubsetIds())
for s in range(1, f1.numberOfSubsets()+1):
    f1.removeFromSubset(s, s-1)
f1.addToSubset(1,1)
f1.addToSubset(1,3)
f1.addToSubset(1,4)
f1.addToSubset(2,2)
f1.addToSubset(2,3)
f1.addToSubset(3,1)
f1.addToSubset(3,2)

print("f1 1: ",f1.getMembers(1))
print("f1 2: ",f1.getMembers(2))
print("f1 3: ",f1.getMembers(3))

f2= nk.Cover(n=5)
f2.allToSingletons()
for s in range(1, f2.numberOfSubsets()+1):
    f2.removeFromSubset(s, s-1)
f2.addToSubset(1,1)
f2.addToSubset(1,2)
f2.addToSubset(1,3)
f2.addToSubset(2,3)
f2.addToSubset(2,4)

print("f2 1: ", f2.getMembers(1))
print("f2 2: ", f2.getMembers(2))

nk.community.OverlappingNMIDistance(normalization=nk.community.Normalization.Max).getDissimilarity(G,f1,f2)

Output:

{1, 2, 3, 4, 5}
f1 1:  {1, 3, 4}
f1 2:  {2, 3}
f1 3:  {1, 2}
f2 1:  {1, 2, 3}
f2 2:  {3, 4}

0.6395516101947395

Note that cover subset IDs start with 1, while the node IDs start with 0.

[fabratu]

Concerning your second question:

There is indeed a bug in the code - getValues() is the culprit here. As a workaround, you can run the algorithm and iterate over the nodes in the graphs to extract the values.

sim = nk.community.CoverF1Similarity(G,f1,f2)
sim.run()
for i in G.iterNodes():
    print(sim.getValue(i))

Output:

0.0
0.8
0.8
0.8
0.0

[fabratu]

Concerning your third question:

You can use the PartitionWriter/ CoverWriter (or their respective reader) to read or write Covers/Partitions. See here for documentation: https://networkit.github.io/dev-docs/python_api/graphio.html?highlight=cover#networkit.graphio.CoverWriter