To create a discrete-time temporal GNN model, we first need to define a configuration file with the suffix yml to specify the model structures and parameters. Here we use the configuration file :code:`parameters_elliptic_egcn_o.yaml` for egcn_o model as an example:
num_hist_steps: 5 # number of previous steps used for prediction
num_hist_steps_min: 3 # only used if num_hist_steps: None
num_hist_steps_max: 10 # only used if num_hist_steps: None
data_loading_params:
batch_size: 1
num_workers: 6
gcn_parameters:
feats_per_node: 50
feats_per_node_min: 30
feats_per_node_max: 312
layer_1_feats: 256
layer_1_feats_min: 30
layer_1_feats_max: 500
layer_2_feats: None
layer_2_feats_same_as_l1: True
k_top_grcu: 200
num_layers: 2
lstm_l1_layers: 125
lstm_l1_feats: 100
lstm_l1_feats_min: 50
lstm_l1_feats_max: 500
lstm_l2_layers: 1
lstm_l2_feats: 400
lstm_l2_feats_same_as_l1: True
cls_feats: 307
cls_feats_min: 100
cls_feats_max: 700
The configuration file is composed of three parts: :code:`dataset_args`, :code:`train` and :code:`gcn_parameters`. Here are their meanings:
- :code:`dataset_args`: This part specifies some configurations for the dataset used. :code:`data` specifies the name of the dataset. :code:`elliptic_args` contains parameters related to the dataset, including the folder location of the dataset, the name of the data file, and so on.
- :code:`train`: This part specifies the training parameters. :code:`use_cuda` indicates whether cuda is used for computation. :code:`use_logfile` indicates whether log files are used to record running processes. :code:`model` indicates the model name to use. :code:`task` indicates the type of the task. :code:`class_weights` is the class weight, which deals with class imbalance. :code:`use_2_hot_node_feats` and :code:`use_2_hot_node_feats` indicates whether one-hot encoding is used. :code:`save_node_embeddings` indicates whether the node embedding is saved. :code:`train_proportion` and :code:`dev_proportion` indicates the ratio of training set and validation set. :code:`num_epochs` indicates the total number of rounds of training. :code:`steps_accum_gradients` indicates the number of steps for gradient accumulation, which is used to implement gradient accumulation. :code:`learning_rate` indicates learning rate. :code:`negative_mult_training` and :code:`negative_mult_test` indicates the multiple of negative sampling at training and test time. :code:`smart_neg_sampling` indicates whether to use negative-only sampling. :code:`seed` indicates the random number seed. :code:`target_measure` and :code:`target_class` denote the target evaluation metric and target category respectively. :code:`early_stop_patience` is the patience value of early stopping, stopping the training if the performance on the validation set does not improve within a certain number of rounds. :code:`eval_after_epochs` indicates how many rounds the evaluation should be performed. :code:`adj_mat_time_window` indicates the time window of the adjacency matrix. :code:`num_hist_steps` indicates the number of historical steps used for prediction. :code:`data_loading_params` indicates data loading parameters, including batch size and number of worker threads.
- :code:`gcn_parameters`: This part specifies the GCN module.These include the number of features per node, the number of features per layer, and the parameters of the LSTM layer. Notice that there are parameters with the suffixes min and max, which means that the parameter values will be randomly generated between min and max based on the random number seed.
After defining the configuration file, we can firstly read the parameters from the configuration file and create a GNN model that supports partition parallelism and use it for later training:
parser.add_argument('--config_file', default='experiments/parameters_elliptic_egcn_o.yaml', type=argparse.FileType(mode='r'), help='optional, yaml file containing parameters to be used, overrides command line parameters')
Then a :code:`build_model` object is created. If needed, we can adjust the model's parameters by modifying the contents of the configuration file. Here we provide 4 models for discrete-time temporal GNNs:
- :code:`EloveGCN`: The EGCN model proposed in `Evolving graph convolutional networks for dynamic graphs <https://ojs.aaai.org/index.php/AAAI/article/download/5984/5840>`__.
- :code:`DySAT`: The DySAT model proposed in `Deep neural representation learning on dynamic graphs via self-attention networks <https://sci-hub.yncjkj.com/10.1145/3336191.3371845>`__.
- :code:`GCRN`: The GCRN model proposed in `Structured sequence modeling with graph convolutional recurrent networks <https://arxiv.dosf.top/pdf/1612.07659.pdf>`__.
- :code:`TGCN`: The TGCN model proposed in `Tag graph convolutional network for tag-aware recommendation <https://xinxin-me.github.io/papers/TGCN.pdf>`__.
