VirtualSoc

Simulate dynamic social networks — with ground-truth labels and realistic node features — so researchers can test algorithms and theories without accessing real social-network data.

VirtualSoc is the open-source simulation library of the paper Simulation and Augmentation of Social Networks for Building Deep Learning Models. Network topology is driven by latent node preferences (the social DNA, sDNA) which double as ground-truth class labels; features, labels and structure are generated coupled together, the way they are entangled in real social networks.

The simulation suite (web UI + API)

pip install -r requirements.txt
python webapp.py          # open http://127.0.0.1:5000

Set the number of people, preference groups, time steps (snapshots with sDNA preference drift between them) and attributes — hit Simulate — watch live progress — download the dataset. Every download is a ready-to-use research bundle:

file	contents
edges_t{K}.csv	undirected edge list of snapshot K (dynamic networks)
labels.csv	ground-truth node labels (sDNA groups)
features_typed.json	realistic attribute values per node
features_encoded.csv	numeric feature matrix (GCN-ready)
stats.json	per-snapshot statistics incl. per-attribute homophily

The same functionality is scriptable over HTTP — POST /api/simulate, GET /api/jobs/<id> (progress), .../result, .../download — see the docstring in webapp.py for the request format. The web layer never executes user-supplied code and enforces hard parameter limits. All generated data is simulated: it contains no real individuals and is safe to share.

Realistic, typed features

Networks can be simulated with typed real-world attributes compared by type-aware dissimilarities inside the sDNA score (see RealFeatures.py):

type	example	dissimilarity
numeric	age, education	abs. difference / range
binary / categorical	gender, city	match / mismatch
geo	(x, y) position	euclidean distance
multihot	interest sets	1 − Jaccard overlap

from RealFeatures import RealFeatureSchema
from Networks import RandomSocialGraphAdvanced

schema = RealFeatureSchema.default_social()   # age, gender, city, geo,
                                              # education, interests
G = RandomSocialGraphAdvanced(labelSplit=[50, 100, 150, 200],
                              realFeatureSchema=schema,
                              useGPU=False, createInGPUMem=False)

sDNA semantics are unchanged (per-feature prefer-similar/dissimilar and weight, mutation for dynamics, labels), and every node also carries a flat numeric encoding (node.features) so exports and GCN training work as before. A custom feature is one FeatureSpec(name, kind, sampler, ...). Full example: ScriptRealFeaturesNetwork.py.

Library usage

• Abstract-feature simulation (as in the paper): ScriptSingleNetwork.py (single network) and ScriptMultiNetwork.py (batches).
• CuPy/CUDA is optional: without CuPy everything runs on the CPU automatically; install cupy-cuda12x to enable GPU-accelerated scoring.
• Tests: python -m pytest tests (19 tests).

Sample generated datasets: data_sample.zip in this repo and many more on Kaggle. To compute graph statistics for generated networks, the standalone rscript is included (R, with its own dependencies; not part of the library).

The paper

A. Wahid-Ul-Ashraf, M. Budka, K. Musial, Simulation and Augmentation of Social Networks for Building Deep Learning Models, arXiv:1905.09087.

Related: the gravitational link-prediction method by the same authors has its reference implementation at AkandaAshraf/akanda-method — including the GCN augmentation from this paper evaluated on Cora.

This work comes from the author's PhD research, funded by Bournemouth University, supervised by the paper's co-authors.