BotRGCN

BotRGCN: Twitter Bot Detection with Relational Graph Convolutional Networks

---

• authors: Shangbin Feng, Herun Wan, Ningnan Wang, Minnan Luo
• link: https://arxiv.org/abs/2106.13092/
• file structure:

├── twibot_22/
│   ├── raw_data/
│   ├── processed_data/
│   ├── utils.py
│   ├── model.py
│   ├── preprocess_1.py
│   ├── train.py # train BotRGCN model
│   ├── Dataset.py
│   ├── preprocess.py # preprocess the dataset
│   ├── dataset_tool.py
│   └── preprocess_2.py
├── cresci_15/
│   ├── raw_data/
│   ├── processed_data/
│   ├── utils.py
│   ├── model.py
│   ├── preprocess_1.py
│   ├── train.py # train BotRGCN model
│   ├── Dataset.py
│   ├── preprocess.py # preprocess the dataset
│   ├── dataset_tool.py
│   └── preprocess_2.py
├── twibot_20/
│   ├── raw_data/
│   ├── processed_data/
│   ├── utils.py
│   ├── model.py
│   ├── preprocess_1.py
│   ├── train.py # train BotRGCN model
│   ├── Dataset.py
│   ├── preprocess.py # preprocess the dataset
│   ├── dataset_tool.py
│   └── preprocess_2.py
└── readme.md

• implement details:

  There are some changes in user numerical properties & user categorical properties due to the lack of relevant data

  1. numerical properties:

     • original: (dim=6)

       followers + followings + favorites + statuses + active_days + screen_name_length

     • twibot-20/cresci-2015/twibot-22: (dim=5)

       followers + followings + statuses + active_days + screen_name_length

  2. categorical properties:

     • original: (dim=11)

       protected + verified + default_profile_image + geo_enabled + contributors_enabled + is_translator + is_translation_enabled + profile_background_image + profile_user_background_image + has_extended_profile + default_profile

     • twibot-20/twibot-22: (dim=3)

       protected + verified + default_profile_image

     • cresci-2015: (dim=1)

       default_profile_image

How to reproduce:

1. specify the dataset by entering corresponding fold

   • cresci-15 : cd cresci_15/
   • twibot-20 : cd twibot_20/
   • twibot-22 : cd twibot_22/

2. preprocess the dataset by running

   python preprocess.py

3. train BotRGCN model by running:

   python train.py

Result:

dataset		acc	precision	recall	f1
Cresci-2015	mean	0.9652	0.9551	0.9917	0.9730
Cresci-2015	std	0.0071	0.0102	0.0025	0.0053
Twibot-20	mean	0.8575	0.8452	0.9019	0.8725
Twibot-20	std	0.0068	0.0054	0.0172	0.0073
Twibot-22	mean	0.7966	0.7481	0.4680	0.5750
Twibot-22	std	0.0014	0.0222	0.0276	0.0142

baseline	acc on Twibot-22	f1 on Twibot-22	type	tags
BotRGCN	0.7966	0.5750	F T G	BotRGCN