The TwitterMancer Project

Code and dataset for the paper:
TwitterMancer:Predicting Interactions on Twitter Accurately
Konstantinos Sotiropoulos¹, John W. Byers¹, Polyvios Pratikakis², and Charalampos E. Tsourakakis^1
1Boston University, USA ²University of Crete, Greece

Dataset

Whole dataset is in the "data/" folder, in a compressed format and split in different files in order for files to be <50MB per github rules. Dataset is anonymized. Alternatively, you can download it from the following link: https://drive.google.com/open?id=16uRqh5Gg1zkV2A-Gi-wx-0EW8FIkjSyn

In order to reproduce results...

Run sequentially:

python create_followSet.py
python construct_features.py {start_date} {end_date}
python prediction.py {start_date} {end_date} > results/prediction_task.txt
python degree_precision.py {start_data} {end_date}

where start_date and end_date are arguments that define the time window we want to use in our dataset
e.g. for using the whole dataset feb 1- feb 28 we have to run:

python create_followSet.py
python construct_features.py 1 28 
python prediction.py 1 28 > results/prediction_task.txt
python degree_precision.py 1 28

Read results from jupyter notebook

We have created a jupyter notebook (called read_results.ipynb ), which reads the output of the prediction.py script (which is saved in a txt file under "results/prediction.txt") and a pickle file, where we have saved the prediction accuracy per embeddedness results and reproduces the main figures and plots from our paper.

Important!

1. Triangles were listed using MACE package Our scripts require a MACE executable inside a "mace/" folder.
2. code is written in Python2.7 and the scikit-learn version used in the experiments is 0.20.2.
3. Dataset is anonymized, so given user IDs do not represent real twitter users.