This project is written in support to the Medium post published here. It is created as a part of the Udacity Data Scientist course, but it can be used as an example of user identification through session data analysis.
The goal of the project is to differentiate Alice (an attacker) from Bob (all the other users) by analysing what sites and when she visits. The analysis is performed The task itself and the data used are taken from this Kaggle competition. While this is not the final version I used to submit my solution to the competition, it incorporates the most important tools to be used in user identification. It is also enough to beat the first baseline: ROC AUC above 0.95343.
I hope you find the project and it's accompanying article useful in learning about bag-of-sites and time series cross-validation.
- A full Anaconda installation
- hyperopt - install via
pip install hyperopt
Important: at the moment of writing this README, hyperopt has issues with one
of the packages it depends upon. If you see model 'bson' has no attribute 'BSON',
then you can circumvent the error this way:
from hyperopt import base
base.have_bson = False
This shouldn't be necessary after the bson package gets updated.
No installation is necessary. Just launch Jupyter Notebook from the root folder.
You can run the explarotory data analysis by running visual_exploration.ipynb.
You can run the mechine learning algorithm by running main.ipynb.
The data for this repo, stored in the data folder, is included as per the terms
of the license for this Kaggle competition.
The project itself is licensed under the GNU Public License (included).