CED : Credible Early Detection of Social Media Rumors

The experiment code of CED, implemented in Python 2.7 and Tensorflow 1.3.0. Due to the large INPUT gap between the proposed model and baseline models, we organize each model into a separate .py file. Later we will modularize the shared code between these models and reconstruct the whole network.

Models:

• 1_CNN_OM: CNN just to deal with original microblogs.
• 2_TF_IDF: SVM classifier using TF-IDF representation vector from 10_parted_posts_seqvec.txt, which has already batched N=10 consecutive reposts together.
• 3_GRU_2: 2-layer GRU to deal with repost sequences.
• 4_CAMI: Our self-implementation model of CAMI.
• 5_(1/2/3)_CED: Our proposed model CED,CED-OM and CED-CNN.

Input Files:

• class_8050.json: Class label and repost feature length of file_name. All 8050 samples are from Rumdect and our published dataset Chinese_Rumor_Dataset.

  {"file_name1":{"class":[0,1], "len":5}, "file_name2":{"class":[1, 0], "len":16}, ......}

• msg_id.json/txt: Padded word embedding ID of original message.

  {"file_name1":[15029,4890,2332,3380,382,6019,320,8524,671,0], "file_name2":[2003,60,1390,0,0,0,0,0,0,0], ......}

• post_id.json: Padded word embedding ID of repost message.

  {"file_name1":[[22,31,1866,468,1170,469,220,5285, ...], [1102,1712,1304,930,127,1712,193,22, ...], ...], "file_name2":[[...], [...], ...].shape = [Padded length of N reposts' words, Corresponding "len" in class_8050.json], ......}

• 10_parted_posts_seqvec.txt: Padded TF-IDF features only for CAMI. Still N=10 to compare with other models.

Large file 10_parted_posts_seqvec.txt and post_id.json can be downloaded here with ks8c.

Train

$ # python model_name_to_train.py , for example:
$ python 5_3_CED_CNN.py

We also show our experiment environment in requirements.txt file.

Citation

If you use this code for research, please cite our paper as follows:

@article{song2019ced,
  title={CED: credible early detection of social media rumors},
  author={Song, Changhe and Yang, Cheng and Chen, Huimin and Tu, Cunchao and Liu, Zhiyuan and Sun, Maosong},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2019},
  publisher={IEEE}
}

Contact

If you have any problem, please feel free to contact us through this email([email protected]).