susheels / tgrank

TGRank

Dataset Links

Reddit - http://snap.stanford.edu/jodie/reddit.csv

Wikipedia - http://snap.stanford.edu/jodie/wikipedia.csv

MOOC - http://snap.stanford.edu/jodie/mooc.csv

LastFM - http://snap.stanford.edu/jodie/lastfm.csv

Preprocessed Enron and UCI datasets are taken from - https://github.com/snap-stanford/CAW

If one wants to use raw datasets use the preprocessing code given in preprocess_data.py and use it as,

python preprocess_data.py --data_dir "your raw data dir" --data "name""

usage: Interface for data preprocessing [-h] [--data_dir DATA_DIR] [--data DATA] [--bipartite]
                                        [--num_node_feats NUM_NODE_FEATS]

optional arguments:
  -h, --help            show this help message and exit
  --data_dir DATA_DIR   Dataset Directory
  --data DATA           Dataset name (eg. reddit, wikipedia, mooc, lastfm)
  --bipartite           Whether the graph is bipartite
  --num_node_feats NUM_NODE_FEATS
                        Number of random node features

Usage

For all parameters of training TGRank follow the commands

python tgrank_listwise_train.py -h

usage: TGRank Interaction Ranking Listwise Training [-h] --data_dir DATA_DIR [--data DATA] [--prefix PREFIX] [--train_batch_size TRAIN_BATCH_SIZE]
                                                    [--eval_batch_size EVAL_BATCH_SIZE] [--num_epochs NUM_EPOCHS] [--num_layers NUM_LAYERS] [--lr LR]
                                                    [--emb_dim EMB_DIM] [--time_dim TIME_DIM] [--num_temporal_hops NUM_TEMPORAL_HOPS] [--num_neighbors NUM_NEIGHBORS]
                                                    [--uniform_sampling] [--patience PATIENCE] [--log_dir LOG_DIR] [--saved_models_dir SAVED_MODELS_DIR]
                                                    [--saved_checkpoints_dir SAVED_CHECKPOINTS_DIR] [--verbose VERBOSE] [--seed SEED]
                                                    [--num_temporal_hops_eval NUM_TEMPORAL_HOPS_EVAL] [--num_neighbors_eval NUM_NEIGHBORS_EVAL]
                                                    [--no_fourier_time_encoding] [--coalesce_edges_and_time] [--train_randomize_timestamps] [--no_id_label]

optional arguments:
  -h, --help            show this help message and exit
  --data_dir DATA_DIR   Dataset directory
  --data DATA           Dataset name (eg. reddit, wikipedia, mooc, lastfm, enron, uci)
  --prefix PREFIX       Prefix to name the checkpoints and models
  --train_batch_size TRAIN_BATCH_SIZE
                        Train batch size
  --eval_batch_size EVAL_BATCH_SIZE
                        Evaluation batch size (should experiment to make it as big as possible (based on available GPU memory))
  --num_epochs NUM_EPOCHS
                        Number of training epochs
  --num_layers NUM_LAYERS
                        Number of layers
  --lr LR
  --emb_dim EMB_DIM     Embedding dimension size
  --time_dim TIME_DIM   Time Embedding dimension size. Give 0 if no time encoding is not to be used
  --num_temporal_hops NUM_TEMPORAL_HOPS
                        No. of temporal hops for sampling candidates during training.
  --num_neighbors NUM_NEIGHBORS
                        No. of neighbors to sample for each candidate node at each temporal hop. This is also the same parameter that samples edges.
  --uniform_sampling    Whether to use uniform sampling for temporal neighbors. Default is most recent sampling.
  --patience PATIENCE   Patience for early stopping
  --log_dir LOG_DIR     directory for storing logs.
  --saved_models_dir SAVED_MODELS_DIR
                        directory for saved models.
  --saved_checkpoints_dir SAVED_CHECKPOINTS_DIR
                        directory for saved checkpoints.
  --verbose VERBOSE     Verbosity 0/1 for tqdm
  --seed SEED           deterministic seed for training. this is different from that by used neighbor finder which uses a local random state
  --num_temporal_hops_eval NUM_TEMPORAL_HOPS_EVAL
                        No. of temporal hops for sampling candidates during evaluation.
  --num_neighbors_eval NUM_NEIGHBORS_EVAL
                        No. of neighbors to sample for each candidate node at each temporal hop during evaluation. This is also the same parameter that samples edges.
  --no_fourier_time_encoding
                        Whether to not use fourier time encoding
  --coalesce_edges_and_time
                        Whether to coalesce edges and time. make sure no_fourier_time_encoding is set and time_dim is 1. else will raise error
  --train_randomize_timestamps
                        Whether to randomize train timestamps i.e. after sampling and before going into TSAR
  --no_id_label         Whether to not use identity label to distinguish source from destinations. Value used to set label diffusion

Example commands

All default parameters are given in the help command.

python tgrank_listwise_train.py --data_dir "your data dir" --data enron --prefix tgrank-listwise --verbose 1
python tgrank_listwise_train.py --data_dir "your data dir" --data wikipedia --prefix tgrank-listwise --verbose 1
python tgrank_listwise_train.py --data_dir "your data dir" --data mooc --prefix tgrank-listwise --verbose 1

Examples of running ablations

Here running wikipedia without source specific label diffusion

python tgrank_listwise_train.py --data_dir "your data dir" --data wikipedia --prefix tgrank-listwise --verbose 1 --no_id_label

Similarly, one can use different parameters like:

--coalesce_edges_and_time,

--train_randomize_timestamps,

--no_fourier_time_encoding