Paper version

This release contains the datasets and the best model reported in the paper

Ablation studies

TODO

Seems like all the results are better than DrBC baseline

Use PReLU in the model
Use Adam optimizer with a big learning rate
Try to have a variable number of edges in the generated graphs
Try dropping edges while training
Graphs are only of 'powerlaw' type.
Use unique convolutions.
Use blocks of convolutions followed with max-pooling and skip connections
Use gradient clipping

Powerlaw graphs with probabilistic m

This release includes a model without RNNs or attention.
Only Graph Convolution layers.

Obtained results on Real datasets (Model was run on 64GB CPU machine with 8vCPUS):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	78.2	76.9	79.1	60.0	TODO
Amazon	90.8	88.3	86.2	78.1	TODO
Dblp	79.7	72.8	75.4	73.4	TODO
cit-Patents	47.7	56.5	63.0	72.4	TODO
com-lj	TODO	TODO	TODO	TODO	TODO

DrBC paper results on Real datasets (Model was run on an 80-core server with 512GB memory, and 8 16GB Tesla V100 GPUs. Trained on GPU, tested on only CPUs):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	73.6	66.7	69.5	57.3	402.9
Amazon	86.2	79.7	76.9	69.3	449.8
Dblp	78.9	72.0	72.5	71.9	566.7
cit-Patents	48.3	57.5	64.1	72.6	744.1
com-lj	67.2	72.6	74.8	71.3	2274.2

Drop Edge with gradient clipping and smaller model

com-youtube

gcn cycles: (4, 4, 6, 6, 8, 8)
conv sizes: (48, 48, 32, 32, 24, 24)
drops: (0.3, 0.3, 0.2, 0.2, 0.1, 0.1)
Largest Linear: 224 x 32
Graph: Data(edge_index=[2, 5975248], x=[1134890, 1], y=[1134890, 1])
{'val_top_1p': 79.10645047585477,
 'val_top_5p': 77.05484280276329,
 'val_top_10p': 78.77767889398972,
 'val_kendal': 59.79420296905783,
 'val_mse': 448.61683670491743,
 'val_max_error': 2120.5370786637122,
 'run_time': 240.21285891532898}

dblp

gcn cycles: (4, 4, 6, 6, 8, 8)
conv sizes: (48, 48, 32, 32, 24, 24)
drops: (0.3, 0.3, 0.2, 0.2, 0.1, 0.1)
Largest Linear: 224 x 32
Graph: Data(edge_index=[2, 17298012], x=[4000148, 1], y=[4000148, 1])
{'val_top_1p': 78.03554911127222,
 'val_top_5p': 70.12404565840195,
 'val_top_10p': 71.69299074532391,
 'val_kendal': 72.43665323132845,
 'val_mse': 370.7392910465112,
 'val_max_error': 471.159696087762,
 'run_time': 982.3556971549988}

Deeper DropEdge Network

Obtained results on Real datasets (Model was run on 64GB CPU machine with 8vCPUS):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	77.2	76.8	79.7	60.1	???
Amazon	91.4	89.5	87.3	78.5	???
Dblp	78.4	69.7	71.3	72.4	???
cit-Patents	Doesn't	have the	dataset	in the	provided URL
com-lj	TODO	TODO	TODO	TODO	TODO

DrBC paper results on Real datasets (Model was run on an 80-core server with 512GB memory, and 8 16GB Tesla V100 GPUs. Trained on GPU, tested on only CPUs):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	73.6	66.7	69.5	57.3	402.9
Amazon	86.2	79.7	76.9	69.3	449.8
Dblp	78.9	72.0	72.5	71.9	566.7
cit-Patents	48.3	57.5	64.1	72.6	744.1
com-lj	67.2	72.6	74.8	71.3	2274.2

Deep GCN with progressive DropEdge

This release includes a model without RNNs or attention.
Only Graph Convolution layers.

Obtained results on Real datasets (Model was run on 64GB CPU machine with 8vCPUS):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	75.7	74.4	77.4	59.9	??
Amazon	91.5	88.5	86.2	78.6	??
Dblp	77.9	71.1	73.7	73.2	??
cit-Patents	Doesn't	have the	dataset	in the	provided URL
com-lj	TODO	TODO	TODO	TODO	??

TODO: get the cit-Patents dataset matching the paper description.

Previous (bigger model, no drop-edge) results on Real datasets (Model was run on 64GB CPU machine with 8vCPUS):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	77.2	76.4	79.0	59.9	198.0
Amazon	92.2	89.4	87.1	78.3	412.0
Dblp	77.3	69.8	71.3	72.3	660.4
cit-Patents	Doesn't	have the	dataset	in the	provided URL
com-lj	69.3	75.9	78.5	71.7	2356.1

DrBC paper results on Real datasets (Model was run on an 80-core server with 512GB memory, and 8 16GB Tesla V100 GPUs. Trained on GPU, tested on only CPUs):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	73.6	66.7	69.5	57.3	402.9
Amazon	86.2	79.7	76.9	69.3	449.8
Dblp	78.9	72.0	72.5	71.9	566.7
cit-Patents	48.3	57.5	64.1	72.6	744.1
com-lj	67.2	72.6	74.8	71.3	2274.2

We need to go deeper with GCNs

This release includes a model without RNNs or attention.
Only Graph Convolution layers.

Obtained results on Real datasets (Model was run on 64GB CPU machine with 8vCPUS):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	77.2	76.4	79.0	59.9	198.0
Amazon	92.2	89.4	87.1	78.3	412.0
Dblp	77.3	69.8	71.3	72.3	660.4
cit-Patents	Doesn't	have the	dataset	in the	provided URL
com-lj	69.3	75.9	78.5	71.7	2356.1

DrBC paper results on Real datasets (Model was run on an 80-core server with 512GB memory, and 8 16GB Tesla V100 GPUs. Trained on GPU, tested on only CPUs):

Dataset	Top 1%	Top 5%	Top 10%	Kendal Tau	Running time
com-youtube	73.6	66.7	69.5	57.3	402.9
Amazon	86.2	79.7	76.9	69.3	449.8
Dblp	78.9	72.0	72.5	71.9	566.7
cit-Patents	48.3	57.5	64.1	72.6	744.1
com-lj	67.2	72.6	74.8	71.3	2274.2

Matching the Vanilla DrBC implementation

This release includes the vanilla DrBC model with the datasets used to evaluate it.

The attached files contain:

• drbc.ckpt contains the model file (PyTorch-lightning checkpoint).
• real (for real-world datasets) used in the paper for model evaluation
• synthetic (for synthetic datasets) used in the paper for model evaluation.

The results can be a bit different from the original DrBC paper due to using different weight initialization, random seed, and learning rate reduction on a plateau. The model is evaluated on a CPU machine with a 3.3 GHz Dual-Core Intel Core i7 Processor and a 16GB RAM (2133 MHz LPDDR3).

Obtained results on Real datasets (com-youtube):

'run_time': 262.6896481513977,
 'val_kendal': 59.04372096009873,
 'val_max_error': 79.49854594229129,
 'val_mse': 1712.2153919586597,
 'val_top_1%': 64.17870990482905,
 'val_top_5%': 70.18186944875228,
 'val_top_10%': 75.83642467551921,