Getting started

Browse, download and cache network datasets from Python.

Seven sources are indexed, each in its own subpackage:

Source	Datasets	Strength
netgamut.snap	109	large, well-known social / web / technological networks
netgamut.netzschleuder	286	domain-tagged, incl. biological, transport and multilayer
netgamut.konect	300	one fixed vocabulary of format, edge type and loops; signed and temporal
netgamut.ogb	12	graph-learning benchmarks, with named node categories
netgamut.pyg	99	the graph-learning corpus: citation, web, social, heterophilous
netgamut.paths	3	curated path data (flight itineraries, navigation trails) for memory networks
netgamut.ahorn	83	higher-order data: hypergraphs, simplicial complexes, revision-pinned

Metadata for all seven ships inside the package, so you can explore every dataset — domains, sizes, statistics, citations — without downloading anything. When you do ask for data it lands in a shared user-level cache, so a dataset is fetched once per machine no matter how many projects, notebooks or virtualenvs use it.

pip install netgamut

For a runnable tour — browsing, loading, and building memory networks from path data — see the netgamut tutorial. Upgrading from 0.2? Generators now return network objects rather than tuples, and some downloads load with different weights; CHANGELOG.md has the call-by-call migration.

One domain vocabulary across every source

The sites label things differently — SNAP groups by provenance (“Autonomous systems graphs”), Netzschleuder tags by domain, AHORN by a vocabulary of its own, KONECT by kind of relationship, OGB and PyG by nothing at all — so all seven are mapped onto one vocabulary. That makes it possible to compare methods across kinds of system rather than kinds of provenance:

import netgamut

netgamut.domains()

{'Social': 366, 'Informational': 159, 'Biological': 51,
 'Technological': 32, 'Transportation': 42, 'Economic': 91}

for match in netgamut.search(domain="Biological", max_nodes=500):
    print(match)
# netzschleuder/blumenau_drug [Biological] (75 nodes, 181 edges)
# netzschleuder/celegans_2019 [Biological] (453 nodes, 4,879 edges)
# netzschleuder/celegans_metabolic [Biological] (453 nodes, 4,596 edges)
# ...

net = netgamut.load("celegans_metabolic")

The kind of network filters across sources too: weighted, bipartite, temporal, multilayer, signed and multigraph. Each source answers from its own catalog, and a dataset whose source cannot say is left out of a True and a False alike:

netgamut.search(multilayer=True)  # Netzschleuder's multiplex networks
netgamut.search(temporal=True, weighted=True, max_edges=100_000)

unknown=True keeps them instead, for every filter their source cannot answer, sizes and node attributes included:

netgamut.search(weighted=False, unknown=True)  # all not known to be weighted

Note Biological: 51 — 36 from Netzschleuder, 5 each from KONECT and OGB, 4 from AHORN, 1 from PyG, and none at all from SNAP. That imbalance is why the other sources exist. The mapping is deliberately visible in netgamut/_core/domains.py rather than buried per source, so it can be argued with in one place.

For anything source-specific, use that source’s own search() (see The sources):

from netgamut import netzschleuder as nz
from netgamut import snap

nz.search(multilayer=True)  # 17 multiplex networks; SNAP has none
nz.search(subtype="Food web")
snap.search(ground_truth=True)  # reference partitions

Find datasets that fit

# Undirected networks small enough to hold in memory
small = snap.search(directed=False, max_edges=1_000_000)

# Everything in one category
snap.datasets("Collaboration networks")

# Free-text over names and descriptions
snap.search("bitcoin")

Each result is a Dataset with .nodes, .edges, .directed, .description, .statistics, .files and .citations.

Categories report different columns — temporal networks have Temporal Edges and Static Edges, review datasets have Number of items — so .columns preserves each table’s own headers verbatim rather than flattening them into a schema SNAP never used.

Download once, reuse everywhere

net = snap.load("ca-GrQc")  # downloads on first use
net = snap.load("ca-GrQc")  # instant, from cache

netgamut.snap.load() returns a netgamut.Network handle, not a parsed graph — so the same call works for a 14k-edge collaboration network and a billion-edge social graph.

net.path  # decompressed file on disk, ready for any tool
net.dataset  # the metadata from above
net.header  # SNAP's own descriptive comment lines
net.size_bytes

for source, target in net.edges(cast=int):  # streamed, flat memory
    ...

for fields in net.rows():  # full records: timestamps, signs, weights
    ...

Conversions are lazy and optional:

graph = net.to_networkx()  # pip install "netgamut[networkx]"
im = net.to_infomap()  # pip install "netgamut[infomap]"
im.run()

Both default to the direction SNAP reports for the dataset.