Datasets and machine learning

Public temporal networks

load_dataset loads public temporal networks from SNAP and SocioPatterns.

• Download and cache. On first use the file is downloaded from the original source and cached in dataset_dir. Set the environment variable TEMPORALGRAPHS_DATA to change the cache directory.
• No redistribution. The data are not redistributed. Every entry of temporal_datasets records the license of the data and the paper to cite, and both are printed when the file is downloaded.

using TemporalGraphs
temporal_datasets()                       # the catalog
g = load_dataset("CollegeMsg")            # 1,899 nodes, 59,835 edges
h = load_dataset("sp-workplace-2015")     # face-to-face contacts, both directions
snapshots(h)                              # 20 s snapshots

Node ids are renumbered 1:n, and original_ids gives the ids of the source. Every edge gets the transition time transition_time, which defaults to 1. SocioPatterns contacts are undirected, so they are loaded in both directions.

The SocioPatterns data are licensed CC0 or CC BY-NC-SA (non-commercial use); see the license and citation fields. The node counts and edge counts of all datasets, except the largest Stack Overflow and Wikipedia ones, were checked against the numbers published by the sources.

GraphNeuralNetworks.jl

After using GNNGraphs (or GraphNeuralNetworks), the conversions to and from GNNGraphs.jl are available.

• Continuous time. GNNGraph(g) is the multigraph of all temporal edges, with their times as edge features t and tt. OrderedEdgeList(h::GNNGraph) converts back.
• Discrete time. TemporalSnapshotsGNNGraph(g; resolution) returns one GNNGraph per snapshot, with the times in tgdata.times. OrderedEdgeList(tg::TemporalSnapshotsGNNGraph) converts back.
• Event graphs. augmented_event_gnngraph is the augmented event graph of Heeg et al. (see Temporal graph isomorphisms), with one-hot node features. A directed GNN on this graph is as expressive as the temporal Weisfeiler–Leman test.

julia> using TemporalGraphs, GNNGraphs

julia> g = OrderedEdgeList(4, [(1, 2, 0, 1), (2, 3, 0, 1), (1, 3, 5, 1), (2, 1, 7, 1)]);

julia> h = GNNGraph(g);

julia> h.num_edges, h.edata.t
(4, [0, 0, 5, 7])

julia> edges(OrderedEdgeList(h)) == edges(g)
true

julia> tg = TemporalSnapshotsGNNGraph(g; resolution = 5);

julia> tg.num_snapshots, tg.num_edges, tg.tgdata.times
(2, [2, 2], [0, 5])

julia> a = augmented_event_gnngraph(g; β = 2);

julia> a.num_nodes, size(a.ndata.x)
(8, (2, 8))

MLDatasets.jl

After using MLDatasets, the graph containers of MLDatasets.jl can be converted.

• OrderedEdgeList(tg::MLDatasets.TemporalSnapshotsGraph, times = 1:tg.num_snapshots) converts a sequence of snapshots, for example the 1,000 brain networks of TemporalBrains() (a 1.6 GB download). MLDatasets.TemporalSnapshotsGraph(g; resolution) converts the other way.
• OrderedEdgeList(g::MLDatasets.Graph; time = :timestamp) reads a static graph whose edges carry their times as edge data.
• OrderedEdgeList(h::MLDatasets.HeteroGraph, relation; time = :timestamp) converts one relation of a heterogeneous graph. For example, ("user", "rating", "movie") of MovieLens gives a bipartite temporal graph, with the source nodes numbered first.

using TemporalGraphs, MLDatasets, DataFrames, CSV   # MovieLens needs DataFrames and CSV
g = OrderedEdgeList(MovieLens("100k").graphs[1], ("user", "rating", "movie"))
# 2,625 nodes (943 users and 1,682 movies), 100,000 ratings
b = OrderedEdgeList(TemporalBrains()[1])             # 102 nodes, 27 snapshots