graines

Classification of Twitter users using multimodal embeddings

Download the code

git clone https://github.com/medialab/graines.git
cd graines

Install requirements

• create a virtual environment with python 3.8
• activate it
• run pip install -r requirements.txt

Create the ground truth

• Move the non_graines_metadata.csv and graines_metadata.csv files inside the graines repo
• run python create_ground_truth.py
• the ground truth is saved in a csv file : "graines_et_non_graines.csv". The seeds get the label 1 and the non-seeds the label 0.

Create your own embeddings

Have a look at the tfidf_on_descriptions.py file: the matrix should be saved as a name_of_your_embedding_model.npy matrix, and have exactly 411 rows. Alternatively topo_count.py measures the topoogical features of candidates. The vectors corresponding to each user should be in the same order as the users in graines_et_non_graines.csv. You can run python tfidf_on_descriptions.py to get an example of the embedding matrix.

Run the test

python main.py --model name_of_your_embedding_model (without .npy in the name of the model) The results are run 5 times with a different train/test split.

To save a complete report to results_binary_classif.csv, run

python main.py --model name_of_your_embedding_model --report

To try a different classifier, run

python main.py --model name_of_your_embedding_model --classifier SVM_RBF_kernel

To test the code only on difficult cases, run

python main.py --model name_of_your_embedding_model --objective difficult_cases

Push your code and results

The .gitignore file should prevent you from loading the users personnal data or any Twitter data we collected.

git commit -am "name of your commit"
git push