The code has been tested with PyTorch 2.3.0 and Python 3.11.5, simulating and constructing the network, and sampling a subnetwork of 50,000 nodes for training and testing.
Verify the performance impact on a single A100 GPU.
Download the weibo_network.txt and place it in the root path of the current project.
Run load_to_sparse_array_fast.py to load web_network.txt and generate weibo.sparse.pl in the current path. The *.sparse.pl file contains the entire social network data of weibo_network.
Run diffusion_for_train_v3_weibo.py to generate a sampled graph of Weibo. The default parameters will load weibo.sparse.pl.
Line 110~118 code in diffusion_for_train_v3_weibo.py define the dataset name, diffusion model, and seed rate of the trainable sub-graph. You can change these parameters to load different datasets using various diffusion models and seed rates. In the seed rate parameter, 1 represents 0.01, 5 represents 0.05, 10 represents 0.1, and 20 represents 0.2 relative to the entire graph node. The default parameters will use weibo dataset, IC diffusion model and 0.01 seed rate to build a trainable sub-graph.
Line 129 code adj = get_top_nodes_and_edge(adj, 50000) in diffusion_for_train_V3_weibo.py decides the size of the sampled graph. 50000means the code will sample a graph containing 50000 nodes and corresponding edges.
The final product of diffusion_for_train_v3_weibo.py is a trainable graph dataset file named *.SG.new, the final product called weibo_mean_IC10.SG.new if you haven't change any parameters.
Run diffusion_for_test.py to train model and verify the impact scope, the result of diffusion_for_test.py represents, on average, how many nodes will be affected by one node. Same as step 2, line 35~43 define the name of file to be read. The default parameters will load a file called weibo_mean_IC10.SG.new and print the result on the screen.
We use smallest dataset jazz and biggest dataset power_grid in DeepIM official datasets to conduct a series experiments.
| our result | paper result | official code result | |
|---|---|---|---|
| LT diffusion model | 39.89 | 99.1 | 81.81 |
| IC diffusion model | 50.65 | 49.9 | 47.82 |
| SIS diffusion model | 71.11 | 74.1 | 69.34 |
The above table is jazz (seed rate = 20%) result, the overall results are consistent with offical result expect LT diffusion model, which has the lowest result. So we use power_grid (seed rate = 20%) dataset and focus on IC and SIS diffusion model to conduct further experiments.
| our result | paper result | official code result | |
|---|---|---|---|
| IC diffusion model | 53.73 | 52.4 | 50.38 |
| SIS diffusion model | 22.04 | 23.8 | 23.81 |
The above table is power_grid (seed rate = 20%) result, the result are still consistent with the official result, so we conduct further experiments on weibo network. We sample two sub-graph by obtaining the first N nodes in original 178w+ weibo graph ourselves, N is 10000 and 50000 respectively.
| weibo(seed rate=20%) 1w node |
weibo(seed rate=1%) 1w node |
weibo(seed rate=1%) 5w node |
|
|---|---|---|---|
| LT diffusion model | 56.1 | / | / |
| IC diffusion model | 37.25 | 14.91 | 36.24 |
| SIS diffusion model | 28.14 | / | / |