calvin-zcx / ndcn

Neural Dynamics on Complex Networks

Neural Dynamics on Complex Networks

Please refer to our paper:

Zang, Chengxi, and Fei Wang. "Neural dynamics on complex networks." In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 892-902. 2020.

@inproceedings{zang2020neural,
  title={Neural dynamics on complex networks},
  author={Zang, Chengxi and Wang, Fei},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={892--902},
  year={2020}
}

Install libs:

conda create --name ndcn 
conda activate ndcn
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch 
conda install networkx 
conda install matplotlib 
conda install scipy
conda install scikit-learn
conda install pandas

Besides, we adapted torchdiffeq in https://github.com/rtqichen/torchdiffeq into a local package due the change of its oringinal codes

Learning continous-time or regularly-sampled graph dynamics

NDCN for mutualistic interation dynamics

Python files: mutualistic_dynamics.py

python mutualistic_dynamics.py  --T 5 --network grid --dump --sampled_time irregular --baseline ndcn --viz --gpu -1 --weight_decay 1e-2

--network *** for underlining graph with choices=['grid', 'random', 'power_law', 'small_world', 'community']
--sampled_time ** for irregularlly-sampled graph dynamics or regularly sampled ones with choices=['irregular', 'equal']
--baseline ** chooses any model from choices=['ndcn', 'no_embed', 'no_control', 'no_graph', 'lstm_gnn', 'rnn_gnn', 'gru_gnn']
Please refer to the code for the detailed parameter choices

Similar commands for heat-diffusion dynamics or gene regulatory dynamics

Python files: heat_dynamics.py and gene_dynamics.py

python heat_dynamics.py  --T 5 --network grid --dump --sampled_time irregular --baseline ndcn --viz --gpu -1 --weight_decay 1e-3
python gene_dynamics.py  --T 5 --network grid --dump --sampled_time irregular --baseline ndcn --viz --gpu -1 --weight_decay 1e-4

Refer to Animations in gif folder

Heat Diffusion on a Grid Graph, Ground Truth

Heat Diffusion on a Grid Graph Learned by our NDCN model

3 dynamics (gene, heat, mutualistic dynamics) on 5 graphs (grid', 'random', 'power_law', 'small_world', 'community' graphs) are shown in gif folder, or download our ppt https://drive.google.com/file/d/1KBl-6Oh7BRxcQNQrPeHuKPPI6lndDa5Y and show in full screen to check our compiled animations.

Semisupervised learning on graphs by our continuous-time GNN model:

--iter 100 experiments:

python dgnn.py --dataset cora  --model  differential_gcn --iter 100   --dropout 0 --hidden 256 --T 1.2 --time_tick 16 --epochs 100 --dump --weight_decay 0.024 --no_control --method dopri5 --alpha 0

a showcased results by my laptop for --iter 5 experiments:

Total time: 772.3850s;
results: 83.180% (mean) +/- 0.756% (std), 83.000% (median);
Min_Acc: 82.600%, Max_Acc: 84.500%
{'no_cuda': False, 'fastmode': False, 'seed': -1, 'epochs': 100, 'rtol': 0.1, 'atol': 0.1, 'lr': 0.01, 'weight_decay': 0.024, 'nHiddenLayers': 0, 'hidden': 256, 'dropout': 0.0, 'dataset': 'cora', 'model': 'differential_gcn', 'iter': 5, 'dump': True, 'delta': 1.0, 'sms': False, 'normalize': False, 'Euler': False, 'T': 1.2, 'time_tick': 16, 'no_control': True, 'method': 'dopri5', 'alpha': 0.0, 'cuda': False}

ndcn's Issues

Problems with semi supervised learning tasks

I reproduced your work. In the semi supervised learning task (gdnn. py), I can't get the same results as you, or even far from you (according to the parameters you give)

Could you share the fps value or parameter of your gif images?

Hello,
I'd like to know the parameters of your synthetic data on this gif repo.

For example, the final T value, fps value, and K, etc...

Thank you!