Project-Prometheus

Classic GNN methods (Pytorch Implementation)

Package Requirements

torch
torch_geometric
torchmetrics
scipy
numpy

Universal Parameters

• Data Paramters

--data_path     str     '../data/'   Input data path
--dataset       str     'Cora'       Choose a dataset from {Cora, CiteSeer, PubMed}
--split         str     'full'       The type of dataset split {public, full, random}

• Training Parameters

--seed          int      123         Random seed 
--epoch         int      100         Number of epochs to train
--lr            float    0.01        Initial learning rate
--weight_decay  float    5e-4        Weight decay (L2 norm on parameters)
--layer/--k     int/int  2/10        Number of layers/k-hop propagations
--hidden        int      64          Number of hidden units
--dropout       float    0.5         Dropout rate

NOTE: --layer indicates numbers of (Graph) Neural Networks layers which include parameters; while --k indicates k-hop propagations which do not have parameters

Model Parameters

• ChebNet Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering

• GCN Semi-Supervised Classification with Graph Convolutional Networks

• GraphSAGE Inductive Representation Learning on Large Graphs

--sample        int      10          Number of neighbors to sample (0 means no sampling)
--batch_size    int      128         Number of nodes per batch
---aggregator   str      mean        Choose a aggregator type from {mean, gcn, meanpooling}

• GAT Graph Attention Networks

--batch_size    int      128         Number of nodes per batch
--head          int      8           Number of head attentions

• FastGCN FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling

--batch_size    int      128         Number of nodes per batch
--sample        int      256         Number of neighbors to sample per layer

• SIGN SIGN: Scalable Inception Graph Neural Networks

• SGC Simplifying Graph Convolutional Networks

• JKnet Representation Learning on Graphs with Jumping Knowledge Networks

--mode          str      cat         Mode for layer fusion {cat, max}

• DropEdge DropEdge: Towards Deep Graph Convolutional Networks on Node Classification

--rate          float    0.8         The percent of the preserve edges. If it equals 1, no sampling is done on adj matrix

• PairNorm PairNorm: Tackling Oversmoothing in GNNs

--scale         float    1.0         Row-normalization scale
--mode          str      PN          Mode for PairNorm {None, PN, PN-SI, PN-SCS}

• DAGNN Towards Deeper Graph Neural Networks

• GCNII Simple and Deep Graph Convolutional Networks

--alpha         float    0.1         Alpha hyperparameters
--theta         float    0.5         Theta hyperparameters

• APPNP Predict then Propagate: Graph Neural Networks meet Personalized PageRank

--alpha         float    0.1         Alpha hyperparameters

• GPRGNN Adaptive Universal Generalized PageRank Graph Neural Network

--alpha         float    0.1         Alpha hyperparameters

• FAGCN Beyond Low-frequency Information in Graph Convolutional Networks

--eps           float    0.3         Eps hyperparameters

• BertNet BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation

Examples

cd GCN
python main.py