DiffIM

Code, datasets and online appendix of the paper “DiffIM: Differentiable Influence Minimization with Surrogate Modeling and Continuous Relaxation”

Requirements

Install requirements by:

pip install -r requirements.txt

resources

We provide the datasets and GNN models used in the experiment. For convenience, you can refer to these resources in the code through the following variables.

• train graphs : graph_WC_train, graph_CL_train, graph_ET_train
• test graphs : graph_WC_test, graph_CL_test, graph_ET_test
• train datasets : WC1000, CL1000, ET1000
• test datasets : WC50, CL50, ET50
• trained GCN models (6-layer) : model_WC, model_CL, model_ET

Experiments

You can conduct the following experiments in a single file experiments.ipynb

1. Generate dataset

Generate a taining dataset of size data_num by generate_dataset function in dataset.ipynb

generate_dataset(graph_name, data_num, saving_tag)

By the same way, generate a test dataset.

for example,

generate_dataset(graph_ET_train, 1000, saving_tag='-1000')  # train dataset
generate_dataset(graph_ET_test, 50, saving_tag='-50')       # test dataset

2. Training

Train GCN model by train function in train.ipynb

train(train_dataset_name, test_dataset_name, hyper_params, saving_name, gpu_num):

‘hyper_params’ can be a dictionary of hyperparameters (or None)

• ‘max_epoch’
• ‘lr’
• ‘lr_gamma’
• ‘gnn_latent_dim’

for example,

train(ET1000, ET50, saving_name='ET_2GCN.pt', hyper_params={'lr':0.002, 'lr_gamma':0.999, 'max_epoch':2000, 'gnn_latent_dim':[128,128,128,128,128,128]}, gpu_num=0)

or you can do the hyperparameter tuning by Optuna:

hparam_tuning(train_dataset_name, test_dataset_name, saving_name, gpu_num)

you can evaluate influence estimation quality of GNN models by

evaluate_quality(model_name, gnn_latent_dim, train_dataset_name, gpu_num)

3. Evaluation

run DiffIM (or baseline) algorithms by pipeline function in algorithm_pipeline.ipynb

pipeline(alg_name, dataset_name, del_edge_num, **alg_kwargs):

alg_name and corresponding **alg_kwargs can be

• ‘random’, ‘outdegree’, ‘betweenness’, ‘pagerank’, ‘KED’, ‘MDS’
  • no extra parameter
• ‘BPM’, ‘MBPM’
  • sampling_num = sampling number of bond percolation method (default=10000)
• ‘RIS’
  • eps = error range (default=0.2)
• ‘greedy’
  • simul_num = number of MC simulation for each step.
• ‘DiffIM’, ‘DiffIM++’
  • model_name
  • gnn_latent_dim
  • gpu_num
• ‘DiffIM+’
  • model_name
  • gnn_latent_dim
  • gpu_num
  • epochs, lr, size_coeff, ent_coeff = GNN optimizing parameters
    • for convenience, you can use **default_hyper_params