Preprint

Preprint here

CTBP dynamic network simulation

This repository contains the code for generating dynamic networks. Dynamic networks are networks that grow one node at a time. The new node first samples its random out-degree, i.e. the number of existing nodes it will attach too. In this repository, this is sampled from the empirical out-degree distribution, which is stored as a python dictionary in out_degree_distr.py. The node then attaches to existing nodes, with probabilities defined by a given combinations of the following mechanisms:

• Affine preferential attachment
• Lognormal aging
• Fitness
  • uniform
  • exponential
  • powerlaw (pareto distributed)

Combinations of these mechanisms leads to $2 \times 2 \times 4 = 16$ different possible generative models. However, certain combinations have nice properties (e.g. scale-free) while others have not-so-nice properties (e.g. being explosive in size).

In addition, this repository contains code for:

• calculating network statistics, referred to here as "static features"
• calculating dynamic feature matrices

Simulation method

The simulation method builds a continuous time branching process. This is a tree. The tree is of size $\text{network size} \times \text{sum of out-degrees}$. The tree is then collapsed to form the final network.

Repository structure

The repository is structured as follows:

• simulation.py contains the main simulation.

  • it depends on empirical_distribution_outdegree.py for sampling the empirical outdegree of each vertex, which in turn uses out_degree_distr_OC.pkl as data.
  • generate_dataset.py contains the script that was used to generate the networks in the paper.

• dynamic_feature_extraction.py contains code for loading the dynamic network and calculating the dynamic feature matrices as specified in the paper.

• static_feature_extraction.py contains code for calculating the static features used in the paper.

  • it depends on dynamic_feature_extraction.py to load the dynamic network and get the static final snapshot as a networkx Graph object.
  • some static featuers are computationally heavy, and for large/many networks, one is recommended to use graph-tool, a Python module with an c++ implementation. See calculating_static_featuers_with_graph_tool.py.