A dynamic heterogeneous graph model with causality enhanced node representations.

This is the source code for paper Causality Enhanced Societal Event Forecasting With Heterogeneous Graph Learning appeared in IEEE ICDM22

The code has been successfully tested in the following environment. (For older dgl versions, you may need to modify the code)

• Python 3.7.9
• PyTorch 1.7.0+cu92
• dgl 0.5.2
• Sklearn 0.23.2

The experiments are conducted on four event datasets collected from Integrated Conflict Early Warning System (ICEWS). These events are encoded into 20 main categories (e.g., protest, demand, appeal) using Conflict and Mediation Event Observations (CAMEO) event codes. Please find example datasets in this Google Drive Link. A brief introduction of the data files is as follows:

• dyn_tf_2014-2015_900.pkl A list of dynamic heterogeneous graphs constructed for samples from 2014 to 2015. The number of word nodes does not exceed 900.

• attr_tf_2014-2015_900.pkl Date and target event (label) information for heterogeneous graphs.
• causal_topics_0.01.pkl Evolving and Multi-view Causal Topics. 0.01 means the significance level is 99%.
• word_emb_300.pkl Word embeddings.

Clone this repo.

git clone https://github.com/yuening-lab/HGC
cd HGC

Download the dataset (e.g., THA_w7h7) from the given link and store them in data filder. Or prepare your own dataset in a similar format. The folder structure is as follows:

- HGC
	- data
		- THA_w7h7
		- ...
	- src

Please run following commands for training and testing under the src folder. We take the dataset THA_w7h7 as the example.

Evaluate baseline models (Examples)

EvolveGCN

python train.py --dataset THA_w7h7 --datafiles dyn_tf_2014-2015_900,dyn_tf_2015-2016_900,dyn_tf_2016-2017_900 --horizon 5 --gpu 1 -m evolvegcn --n-hidden 64 --n-layers 1 --note "" --train 0.4 --patience 15

HGT

python train.py --dataset THA_w7h7 --datafiles dyn_tf_2014-2015_900,dyn_tf_2015-2016_900,dyn_tf_2016-2017_900 --horizon 5 --gpu 1 -m temphgt --n-hidden 64 --n-layers 1 --note "" --train 0.4 --patience 15

Evaluate the HGC model

Full model

python train.py --dataset THA_w7h7 --datafiles dyn_tf_2014-2015_900,dyn_tf_2015-2016_900,dyn_tf_2016-2017_900 --horizon 5 --gpu 5 -m hgc --n-hidden 64 --n-layers 1 --note "cau0.05" --train 0.4 --n-topics 50 --causalfiles causal_topics_0.05 --patience 15

Variant model wthout causal

python train.py --dataset THA_w7h7 --datafiles dyn_tf_2014-2015_900,dyn_tf_2015-2016_900,dyn_tf_2016-2017_900 --horizon 5 --gpu 6 -m hgc_no_cau --n-hidden 64 --n-layers 1 --note "" --train 0.4 --n-topics 50  --patience 15

Please cite our paper if you find this code useful for your research:

The reference will be updated soon.