CausalRD

This repo provides a reference implementation of CausalRD: A Causal View of Rumor Detection via Eliminating Popularity and Conformity Biases

Dependencies

python 3.7

pytorch 1.8.1

pytorch_geometric 1.7.0

bert-as-service 1.10.0

Dataset

We use the raw data in debias phase and pre-processed data released by Bi-GCN in inference phase. You can download the raw data and move it to ./CausalRD_debias/data/

Usage

Debias Phase

1. Preprocessing

We first leverage bert-as-service to get the embeddings of source tweets. We use BERT-Base (uncased_L-12_H-768_A-12) as the bert model. Please download it from above link and unzip it to bert/, create bert-as-service server and run the client.

pip install bert-serving-server  # server
pip install bert-serving-client  # client, independent of `bert-serving-server`
bert-serving-start -model_dir ./CasualRD/CausalRD_debias/bert/uncased_L-12_H-768_A-12/ -num_worker=4 
python ./CasualRD/CausalRD_debias/bert/bert_pre.py

2. Structral Negative Sampling

We then generate positive and negative samples by Structral Negative Sampling strategy.

python ./CausalRD/CausalRD_debias/sampling.py

3. Optimizing

python ./CausalRD/CausalRD_debias/optimizing.py --dataset twitter15 --n_layers 7

This will generate a directory ./graph/, move it to ./CausalRD/CausalRD_inference/data/Twitter15

Inference Phase

create "Twitter15graph" folder and "Twitter16graph" folder in the ./CausalRD_inference/data folder

python ./CausalRD/CausalRD_inference/Process/getTwittergraph.py Twitter15
python ./CausalRD/CausalRD_inference/train.py --datasetname Twitter15 --weight_decay 0.0001 --iteration 100 --lr 0.0005