[BUG] is_edge_ids in uniform_sampling does not seem to do anything

[VibhuJawa]

Describe the bug

When we pass is_edge_ids=True we dont seem to do anything with it in the uniform_neighbor_sample function (see below). We should triage this and fix it accordingly.

cugraph/python/cugraph/cugraph/sampling/uniform_neighbor_sample.py

Lines 23 to 50 in 7d8f0fd

	def uniform_neighbor_sample(G,
	start_list,
	fanout_vals,
	with_replacement=True,
	is_edge_ids=False):
	"""
	Does neighborhood sampling, which samples nodes from a graph based on the
	current node's neighbors, with a corresponding fanout value at each hop.
	Note: This is a pylibcugraph-enabled algorithm, which requires that the
	graph was created with legacy_renum_only=True.
	Parameters
	----------
	G : cugraph.Graph
	cuGraph graph, which contains connectivity information as dask cudf
	edge list dataframe
	start_list : list or cudf.Series (int32)
	a list of starting vertices for sampling
	fanout_vals : list (int32)
	List of branching out (fan-out) degrees per starting vertex for each
	hop level.
	with_replacement: bool, optional (default=True)
	Flag to specify if the random sampling is done with replacement

[VibhuJawa]

MRE:
EDIT: Cleaned up MRE to be easier to understand.

import cudf
import cugraph
import numpy as np

df = cudf.DataFrame({'source':[1,1],
                     'destination':[2,2],
                     'eids':[0,1]
                    })
df = df.astype(np.int32)


G = cugraph.Graph(directed=True)
G.from_cudf_edgelist(df,edge_attr='eids')

sample_df = cugraph.uniform_neighbor_sample(
    G,
    cudf.Series([1]),
    fanout_vals=[2],
    is_edge_ids=True
)
print(sample_df)

   indices  sources  destinations
0        0        1             2

Expected behavior:

   indices  sources  destinations
0        0        1             2
1        1        1             2

[alexbarghi-nv]

The C++ code operates only on vertices but I can always automatically get the start vertices from the provided edges. Also shouldn't the expected result be:

   indices  sources  destinations
0        1        1             2

[rlratzel]

While we're addressing this, we should also update the docstring.

[VibhuJawa]

import cudf
import cugraph
import numpy as np

df = cudf.DataFrame({'source':[1,1],
                     'destination':[2,2],
                     'eids':[0,1]
                    })
df = df.astype(np.int32)


G = cugraph.Graph(directed=True)
G.from_cudf_edgelist(df,edge_attr='eids')

sample_df = cugraph.uniform_neighbor_sample(
    G,
    cudf.Series([1]),
    fanout_vals=[2],
    is_edge_ids=True
)
print(sample_df)

Nope, i want both the edges. Index 0 and Index 1. Maybe below example is easier to get my point across ?

The index_ids here are 10 and 11 and I do not renumber to be clearer.

import cudf
import cugraph
import numpy as np

df = cudf.DataFrame({'source':[1,1],
                     'destination':[2,2],
                     'eids':[10,11]
                    })
df = df.astype(np.int32)


G = cugraph.Graph(directed=True)
G.from_cudf_edgelist(df,edge_attr='eids', renumber=False, legacy_renum_only=False)

sample_df = cugraph.uniform_neighbor_sample(
    G,
    cudf.Series([1]).astype(np.int32),
    fanout_vals=[2],
    is_edge_ids=True
)
print(sample_df)

   sources  destinations  indices
0        1             2       10
0        1             2       11

[alexbarghi-nv]

Ok, so you want the current behavior, but with the is_edge_ids flag specifying whether the weights are edge ids rather than always assuming they are, correct?

[alexbarghi-nv]

My interpretation was that if is_edge_ids was true, it should treat the start ids as edge ids and automatically select the vertices from those edges.

[VibhuJawa]

Ok, so you want the current behavior, but with the is_edge_ids flag specifying whether the weights are edge ids rather than always assuming they are, correct?

Yup. Thats my understanding and then treating sources, destinations,edge_ids as unique identifiers so that they are sampled.

[VibhuJawa]

My interpretation was that if is_edge_ids was true, it should treat the start ids as edge ids and automatically select the vertices from those edges.

Okay. Yup, thats a perfectly viable understanding too . I guess we really need a docstring here to clarify what this was supposed to do.

[alexbarghi-nv]

ok, so there are 2 problems here:

1. You need to use cugraph.MultiGraph since cugraph.Graph will automatically drop duplicate edges
2. Apparently it ignores the duplicate edge even when using a cugraph.MultiGraph which is either a bug in plc graph creation or uniform_neighbor_sample itself.