The project explores the performance of various machine learning algorithms for the retweeting prediction problem, depending on the provided features.

Tweets are labeled into two classes:

0: tweets retweeted only once

1: tweets retweeted more than once

Types of features:

a) Content features only, extracted by a transformer model InfoCoV/Cro-CoV-cseBERT

b) Various tabular features representing Twitter users and their interactions

c) Joined features a) and b)

Tested on Python 3.9.9

pip install -r requirements.txt (virtualenv recommended)

Original input data: ./data/original/Org-Retweeted-Vectors_preproc.csv

Create folders for processed data:

./data/intermediary/

./data/prepared/

./notebooks/exploration/features_analysis.ipynb

Run scripts in the preparation folder:

1. bert_extract.py to extract content features
2. tran_val_test.py to create train, validation, and test splits for id_str
3. feature_preparation.py for training, validation, and test df, without content features
4. full_feature_joing.py for training, validation, and test df for joined features and content features alone

Alternatively, you can download all the original and prepared data from here: data.zip

After download, unzip all to data.

Notebooks in the notebooks/baselines folder investigate MLP and Random Forest runs:

1. content_only_2_labels.ipynb, content features only
2. features_only_2_labels, network and tabular features without content features
3. full_features_2_labels, all features

Run train.py "some_config.yaml" by providing one of the configs in the configs folder.

The config file specifies the model type and all related parameters.

The training script runs multiple experiments to find the best hyperparameters by using Optuna optimization.

Results and model checkpoints are stored in the experiments folder.

You can download all results and model checkpoints from here: experiments.zip

After download, unzip all to experiments.

Notebooks in notebooks/results_analysis analyze all the results from experiments.