CACL: Community-Aware Heterogeneous Graph Contrastive Learning for Social Media Bot Detection

Accepted by ACL 2024 Findings

CACL is a Community-Aware Heterogeneous Graph Contrastive Learning framework and we apply it to social media bot detection.

The implementation of CACL is mainly based on Pytorch and Pytorch Geometric API.

Overview

The steps to reimplement this work mainly contain:

• Data preprocessing

  • Three datasets are used and programs for preprocessing can be found as predata_{dataset name}.py.
  • Just download the datasets and change the file paths augments.

• Pretrain

  • Community-aware module is employed to perform community detection and link prediction, which needs to be pretrained to prevent cold-start problem.
  • Simply change the preprocessed data path augment and use pretrain.py.

• Train

  • After obtaining preprocessed data and pretrained community-aware module weight, use train.py to train the full model.
  • Hyperparameters can be adjusted in the function super_parament_initial() in utils.py.

• Test
  After training the model using train.py, the test results will be shown and saved.

Datasets

We have used three datasets throughout the entire work. You may need to contact the author to get access to some of them.

• Cresic-15 [paper] [dataset]

• Twibot-20 [paper] [github]

• Twibot-22 [paper] [github] [dataset]

• Other datasets may be helpful [web path]

Training details

Using dataset Cresci-15 and backbone HGT for example.

Once you preprocess the dataset using predata_cresci15.py, then you can:

• Pretrain the community-aware module

  python3 pretrain.py --dataset cresci15 --basic_model HGT

• Train the CACL with HGT

  python3 pretrain.py --dataset cresci15 --basic_model HGT --max_error_times 5

Here are some key options of the hyperparameters

• basic_model: CACL framework support 3 backbones as the convolutional layer including GAT, SAGE, and HGT.
• num_layer: the layer number of the convolutional network, we use 2 by default.
• lr_warmup_epochs: during the initial training warm-up phase of the model, we increase the weight of the contrastive loss, aiming for the model to quickly find the optimal point.
• max_error_times: we use the validation dataset for early stopping.
• cluster: we implement several cluster method for community detection, we use randomwalk by default.

The details of other optional hyperparameters can be found in the function super_parament_initial() in utils.py

Citation

Please consider citing the following paper when using our code for your application.

@inproceedings{CACL2024,
  author       = {Sirry Chen and
                  Shuo Feng and
                  Songsong Liang and
                  Chen{-}Chen Zong and
                  Jing Li and
                  Piji Li},
  editor       = {Lun{-}Wei Ku and
                  Andre Martins and
                  Vivek Srikumar},
  title        = {{CACL:} Community-Aware Heterogeneous Graph Contrastive Learning for
                  Social Media Bot Detection},
  booktitle    = {Findings of the Association for Computational Linguistics, {ACL} 2024,
                  Bangkok, Thailand and virtual meeting, August 11-16, 2024},
  pages        = {10349--10360},
  publisher    = {Association for Computational Linguistics},
  year         = {2024},
  url          = {https://aclanthology.org/2024.findings-acl.617},
  timestamp    = {Tue, 27 Aug 2024 17:38:11 +0200},
  biburl       = {https://dblp.org/rec/conf/acl/ChenFLZLL24.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}