HIF

Heterogeneous GNN Driven Information Prediction Framework

Description

Our HIF is implemented mainly based on the following libraries (see the README file in source code folder for more details):

• PyTorch: https://pytorch.org/
• PyTorch Lightning: https://www.pytorchlightning.ai/
• DGL: https://www.dgl.ai/
• PytorchNLP: https://github.com/PetrochukM/PyTorch-NLP
• Networkx: https://networkx.org/

File Tree

Our main project file structure and description:

HIF
├─ README.md
├─ environment.yaml       # our environment list
├─ data
│    ├─ processed       # processed data will be made by program
│    ├─ raw       # our raw data
├─ models       # package of models (HIF and its variants)
│    ├─ __init__.py
│    ├─ classfier.py
│    ├─ hetero.py
│    ├─ predictor.py
│    ├─ test.py
├─ nn       # package of neural network layers
│    ├─ coattention
│    ├─ conv
│    ├─ functional
│    ├─ metrics
│    ├─ nlp
│    ├─ tcn
│    ├─ time
│    ├─ __init__.py
│    ├─ conv.py	# graph convolution layer
│    ├─ embedding.py       # embedding module
│    ├─ noise.py
│    ├─ positional_encoding.py
│    ├─ reinforcements.py
│    └─ readout.py	# output layer
├─ run.py	# running entrance
└─ utils	# utils of drawing, dataloading, tensor and parameter propcessing
       ├─ data
       ├─ draw
       ├─ graph_op
       ├─ __init__.py
       ├─ additive_dict.py
       ├─ ckpt.py
       ├─ clean_lightning_logs.py
       ├─ group_tensorborad_data.py
       ├─ indexer.py

Installation

Installation requirements are described in environment.txt

• Create Conda Environment:

  Open a terminal or command prompt and run the following command:

  conda env create -f environment.yml

• Activate Conda Environment:

  After the installation is complete, activate the newly created Conda environment using the following command:

  conda activate <environment_name>

• Verify Environment Installation:

  You can verify whether the environment is correctly installed:

  conda list

Usage

Before it starts running, please make sure that HIF is located in the /root/hif directory

Then, get helps for all parameters of data processing, training and optimization:

python run.py  --help

Run:

python run.py --paramater value

Experiment Settings

Here we list some basic settings of parameters for your reference:

Basic settings

Parameter	Value	Description
gpus	1	Number of GPUs to train on (int) or which GPUs to train on (list or str) applied per node.
time_loss_weig	5e-5	Wights of time nodes similarity.
num_workers	4	Number of workers.
num_heads	16	Number of attention heads.
log_every_n_steps	5	How often to log within steps.
batch_size	64	Batch Size.
name	SSC	Datasets name.
observation	1.0	Observation time.
sample	0.05	Sample.
accumulate_grad_batches	1	Accumulates grads every k batches or as set up in the dict.
in_feats	256	Dimension of inputs.
learning_rate	5e-3	Learning rate to optimize parameters.
weight_decay	5e-3	Weight decay.
l1_weight	0	L1 loss.
patience	35	Patience of early stopping.
num_time_nodes	6	Number of time nodes.
soft_partition	2	Time node soft partition size.
num_gcn_layers	2	The number of gcn layers.
readout	ml	The readout module.
num_readout_layers	3	The number of internal layers of readout module.
time_decay_pos	all	Position of time decay.
time_module	transformer	Time embedding module.
num_time_module_layers	4	The number of layers of time module, such as rnn layers or transformer encoder layers.
dropout	0.3	Dropout.
dropout_edge	0	Dropout edges.
noise_weight	0	Weight of noise.
noise_rate	1	Coverage rate of noise.
noise_dim	1	Dimension of noise.
hop	1	Hops of sampled follower.
alpha	10	Weight to adjust sampled follower
beta	100	Minimum number of sampled follower.
source_base	4	Weight of source nodes.
reposted_base	5	Weight of reposted nodes.
leaf_base	6	Weight of leaf nodes.
method	bfs	Method to sample follower.
gradient_clip_val	1	The value at which to clip gradients.
max_epochs	300	Stop training once this number of epochs is reached.
enable_progress_bar	False	Whether to enable to progress bar by default.
enable_model_summar	False	Whether to enable model summarization by default.