aawil/wiki-vandals

Wikipedia vandalism exploration and detection

Exploration and detection of anonymous vandalism of Wikipedia

My fifth project at Metis. I started with around 730,000 anonymous edits to English Wikipedia extracted using WMF Quarry (query here), and pulled in data from the ORES API, the Wikipedia API, and the MaxMind GeoLite2 IP geocoding database. I also added similar metadata to a 2015 list of human-labeled revisions used to train the ORES damage-probability model.

I built my own version of a vandalism-probability scoring model, trained and validated on the 2019 machine-labeled data, and tested on the 2015 human-labeled data.

I also created a web app for exploring the May 2019 anonymous edits I collected.

A brief guide to the files in this directory:

• get_damage_probs is a script for gathering ORES damage probabilities in 50-revision batches, which are needed for the SQL data preparation.
• p5utils.py includes several utility functions used in the data preparation and modeling notebooks, including the ‘bad words’ regex lists partially swiped from revscoring.
• sql_data_prep_and_EDA.ipynb prepares the data downloaded from the WMF Quarry SQL server for use in the web app and model building.
• labeled_data_prep_and_EDA.ipynb does the same thing for the 20k human-labeled revision set.
• webapp_data_prep.ipynb creates the dataframe pickles used in the web app.
• vandal-app is the bare bones of the app itself: just the app.py, the small api.py, and the index.html template.
• vandal_model.ipynb includes model building and visualization.