- Language-Agnostic Twitter Bot Detection
-
authors: Jurgen Knauth
-
file structure:
├── generate_features.py # convert raw dataset into standard format
└── Adaboost.py # train model on every dataset- implement details:
- We ignored the content-based features due to the lack of the information in our datasets
- We did not use the SMOTEENN method to balance the negative samples, since our datasets are well balanced and adding SMOTEENN did not improve the results
- All datasets except Twibot-20 lack some of the core features in account based features such as default profile and geo-enabled, and some of the users’ data is missing ( i.e. no friends count t, follower count etc. ) and were set to 0 in our experiments
the data has been preprocessed and stored in folders e.g. /Twibot-20
change the dataset name dataset='Twibot-22'
and run Adaboost.py
you can check the results in /results/dataset.log
| dataset | acc | precison | recall | f1 | |
|---|---|---|---|---|---|
| Twibot-22 | mean | 0.7125 | 0.7990 | 0.5214 | 0.3709 |
| Twibot-20 | mean | 0.8191 | 0.9656 | 0.7630 | 0.8524 |
| botometer-feedback-2019 | mean | 0.7597 | 0.3171 | 0.5909 | 0.4127 |
| cresci-rtbust-2019 | mean | 0.5000 | 0.5741 | 0.5124 | 0.5415 |
| cresci-stock-2018 | mean | 0.8874 | 0.9989 | 0.8883 | 0.9403 |
| midterm-2018 | mean | 0.8393 | 0.9991 | 0.8399 | 0.9126 |
| cresci-2017 | mean | 0.9022 | 0.9156 | 0.9535 | 0.9342 |
| gilani-2017 | mean | 0.4967 | 0.3517 | 0.4400 | 0.3910 |
| cresci-2015 | mean | 0.8592 | 0.8570 | 0.9740 | 0.9118 |