Pytorch implementation for Yang Liu, Yi-Fang Brook Wu "Early Detection of Fake News on Social Media Through Propagation Path Classification with Recurrent and Convolutional Networks". AAAI 2018.

Note: We do not publish the dataset used in this code. The authors of this paper also did not publish the dataset

Preprocessing

In the project root directory,

1. Prepare dataset

   python preprocess/prepare_dataset.py --config config.json --max-length 5 --output output_5.json --n_processes 30

2. split train/dev dataset

   We need to split the dataset into train/dev carefully. If the dataset is split into train/dev randomly, the features used in RNN/CNN may be duplicated. This is because a number of users in the dataset emits several tweets over entire dataset time.

   python preprocess/split_train_dev.py --input output_5.json -d 2018-01-01 --train_output train_5.json --dev_output dev_5.json

Train

In the project root directory,

fakenews_detection/main.py --mode train --config config.json --train_file train_5.json --dev_file dev_5.json --lr 0.2 --batch_size 128 --cuda 1