Inf. Max. Simulator for Multilayer Networks under ICM

A repository to generate dataset with marginal efficiency for each actor from the evaluated network and evaluete various influence maximisation methods.

• Authors: Piotr Bródka, Michał Czuba, Adam Piróg, Mateusz Stolarski
• Affiliation: WUST, Wrocław, Lower Silesia, Poland

Configuration of the runtime

First, initialise the enviornment:

conda env create -f env/conda.yaml
conda activate infmax-simulator-icm-mln

Then, pull the submodule with data loaders and install its code:

git submodule init && git submodule update
pip install -e _dataset/infmax_data_utils

A final step is to install wrappers for influence-maximisation methods into the conda environment. We recommend to link it in editable mode, so after you clone particular method just install it with pip install -e ../path/to/infmax/method.

Data

Dataset is stored in a separate repository bounded with this project as a git submodule. Thus, to obtain it you have to pull the data from the DVC remote. In order to access it, please sent a request to get an access via e-mail (michal.czuba@pwr.edu.pl). Then, simply execute in a shell:

• cd _data_set && dvc pull nsl_data_sources/raw/multi_layer_networks/*.dvc && cd ..
• cd _data_set && dvc pull nsl_data_sources/spreading_potentials/multi_layer_networks/*.dvc && cd ..

Structure of the repository

.
├── _configs                -> eample configuration files to trigger the pipeline
├── _data_set               -> networks to compute actors' marginal efficiency for
├── _test_data              -> examplary outputs of the dataset generator used in the E2E test
├── _output                 -> a directory where we recommend to save results
├── env                     -> a definition of the runtime environment             
├── src
│   ├── evaluators          -> scripts to evaluate performance of infmax methods
│   ├── generators          -> scripts to generate SPs according to provided configs
│   └── icm                 -> implementations of the ICM adapted to multilayer networks
├── README.md          
├── run_experiments.py      -> main entrypoint to trigger the pipeline
└── test_reproducibility.py -> E2E test to prove that results can be repeated

Running the pipeline

To run experiments execute: run_experiments.py and provide proper CLI arguments, i.e. a path to the configuration file. See examples in _configs for inspirations. The pipeline has two modes defined under the run:experiment_type field.

Generating dataset

The first one ("generate"), for each evaluated case of ICM, produces a csv file a folllowing data regarding each actor of the network:

actor: int              # actor's id
simulation_length: int  # nb. of simulation steps
exposed: int            # nb. of infected actors
not_exposed: int        # nb. of not infected actors
peak_infected: int      # maximal nb. of infected actors in a single sim. step
peak_iteration: int     # a sim. step when the peak occured

Evaluating seed selection methods

The second option ("evaluate") serves as an evaluation pipeline for various seed selection methods which are defined in the study. That is, for each evaluated case of ICM it produces a following csv:

infmax_model: str       # name of the model used in the evaluation
seed_set: str           # IDs of seed-actors aggr. into str (sep. by ;)
gain: float             # gain obtained using this seed set
simulation_length: int  # nb. of simulation steps
exposed: int            # nb. of active actors at the end of the simulation
not_exposed: int        # nb. of actors that remained inactive
peak_infected: int      # maximal nb. of infected actors in a single sim. step
peak_iteration: int     # a sim. step when the peak occured
expositions_rec: str    # record of new activations aggr. into str (sep. by ;)

GPU acceleration for the computations

Selecting GPU (for a tensor runner) is possible only by setting an env variable before executing the Python code, e.g. export CUDA_VISIBLE_DEVICES=3

For instance:

conda activate infmax-simulator-icm-mln
export CUDA_VISIBLE_DEVICES=2
python generate_sp.py _configs/example_tensor.yaml

Results reproducibility

Results are supposed to be fully reproducable. There is a test for that: test_reproducibility.py.