/Troll-d

Classify message into 2 categories : Troll or not troll

Primary LanguageJavaScript

Troll-d

Troll-d is an application hosted on the google cloud app engine, which predicts if a message can be considered as a "troll" message, with a Random Forest algorithm.

The model

The raw text go through a Bag of Words and a Tf-Idf transformation. A truncated SVD is then applied to reduce the dimensionnality and get a vector of 300 dimensions describing the text. This vector is then given to a trained and optimized Random Forest Algorithm, labeling the message as "troll" or "not troll". An AUC of 0.965 is achieved on the validation set.

The notebook modelStudy.ipynb is a study to find the algorithm and parameters giving the best AUC possible (it should be opened with spyder to have control over the cells).

The app

The model is hosted on a Flask application on the google cloud app engine. The app folder contains all the files hosted on the cloud.

  • main.py : flask application
  • app.yaml : configuration file
  • static : contains the machine learning pipeline and the .css
  • templates : html files