Graph-Distribution-Learning

visualize_data.py contains functions display_graph_stats and display_label_stats which are used to visualize the data generically or to visualize the label distributions of the data. To use either, create the dataset as a geo.data.Data() object, then pass it into the function. display_label_stats will also need an additional string that denotes the dataset's name.

The original LDL implementation mentioned in https://doi.org/10.48550/arXiv.1408.6027 is found in the LDL folder. To run, just change the load dataset on line 18 of iislldDemo.m to which ever dataset you want to to run (acm50, yelp50, yelp2_50, dblp50). Run this file and it will return the results for LDL.

By default, all the gcn-based models(including our model) run under GPU setting. To use CPU mode, modify the device='cuda' to device='cpu' in the utils.py file and add additional args --gpu_id -1 when running the following commands.

grid_search_dynamic.py provides a demo template for hyperparameter search of GLDL dynamic setting. single_run.py provides a demo for a single training case for GLDL.

1. To reproduce the results in the paper, we have trained the model and saved the trained model. The trained model state is in the ./src/trained_model/ directory in a zip format. Please unzip the file and put them under the trained_model directory first.
2. set up the environment following the requirements.txt
3. For different datasets, run the following commands: ACM:
   GCN:

python over_smothing.py --dataset acm --p 50 --hidden_channel 128 --num_layers 3 --patience 10 --lr 0.0001 --weight_decay 0.0005

static:

python quick_run.py --dataset acm --p 50 --lr 0.0001 --weight_decay 0.0005 --gcnhidden 320 --pgehidden 64 --freql 30 --freqv 50 --gcnvlayers 2 --gcnllayers 2 --pgelayers 3 --mode static --epochs 500 --patience 50

dynamic:

python quick_run.py --dataset acm --p 50 --patience 50 --epochs 500 --mode dynamic --inner_epochs 70 --gcnhidden 450 --pgehidden 64 --freql 50 --freqv 30 --gcnvlayers 2 --gcnllayers 2 --pgelayers 3 --lr 0.0001 --weight_decay 0.0005

DBLP
GCN:

python over_smothing.py --dataset dblp --p 50 --hidden_channel 32 --num_layers 2 --patience 10 --lr 0.001 --weight_decay 0.0005

static :

python quick_run.py --dataset dblp --p 50 --mode static --gcnhidden 200 --freql 30 --freqv 80 --gcnvlayers 2 --gcnllayers 1 --lr 0.0001 --weight_decay 0.0005 --epochs 500 --patience 100

dynamic:

python quick_run.py --dataset dblp --p 50 --gcnhidden 300 --pgehidden 64 --freql 30 --freqv 20 --lr 0.0001 --weight_decay 0.0005 --gcnvlayers 2 --gcnllayers 2 --pgelayers 2 --mode dynamic --save 1 --epochs 500 --patience 20

Yelp
GCN:

python over_smothing.py --dataset yelp --p 50 --hidden_channel 32 --num_layers 3 --patience 50 --lr 0.0001 --weight_decay 0.005

static:

python quick_run.py --dataset yelp --p 50 --gcnhidden 64 --gcnvlayers 2 --gcnllayers 2 --freqv 20 --freql 60 --lr 0.0001 --weight_decay 0 --mode static --patience 20 --epochs 500

dynamic:

python quick_run.py --dataset yelp --p 50 --gcnhidden 64 --pgehidden 32 --freql 60 --freqv 20 --lr 0.0001 --weight_decay 0.0005 --gcnvlayers 4 --gcnllayers 2 --pgelayers 2 --mode dynamic --epochs 500 --patience 20 --inner_epochs 50

Yelp2
GCN:

python over_smothing.py --dataset yelp2 --p 50 --hidden_channel 128 --num_layers 3 --patience 50 --lr 0.0001 --weight_decay 0.0005 --max_epochs 1000

static:

python quick_run.py --dataset yelp2 --p 50 --gcnhidden 256 --pgehidden 64 --freql 60 --freqv 60 --lr 0.001 --weight_decay 0.0005 --gcnvlayers 4 --gcnllayers 1 --pgelayers 2 --mode static --epochs 500 --patience 50 --inner_epochs 50

dynamic:

python quick_run.py --dataset yelp2 --p 50 --gcnhidden 320 --pgehidden 64 --freql 30 --freqv 60 --lr 0.0001 --weight_decay 0.0005 --gcnvlayers 3 --gcnllayers 2 --pgelayers 2 --mode dynamic --epochs 500 --patience 50 --inner_epochs 50