Dauren Baitursyn

News crawling aggregator. It crawls the information from the government news' website and saves it to a database.

The project has a search engine to retrieve the articles based on the cosine similarity of the frequency-inverse document frequency (TF-IDF) representation of the queries and articles.

A web-user identification is a hot research topic on the brink of sequential pattern mining and behavioral psychology.

Here we will try to identify a user on the internet by tracking their sequence of attended web pages. The algorithm to be built will take a webpage session—a series of web pages attended consequently by the same person—and predict whether it belongs to Alice or somebody else.

As for 23.12.18, my current LB standing is 87th out of 2,000ish.

For this competition, the data was time-series session information regarding user browser history. We had to distinguish between a regular user and an intruder user based on sites visited and time of visit information.

I performed the exploratory data analysis (EDA), data wrangling, and feature engineering to achieve a ROC-AUC score of 0.95856.

IMPLEMENTED FEATURES
• Dummy hour feature of the session start time (from now on start time).
• Sin and cos transformation of the start time.
• Active start hours of the intruder.
• Dummy weekday feature of the start time.
• Active weekday of intruder feature.
• Dummy month feature of the start time.
• Sin and cos transformation of the year and day feature.
• Session length.
• Session sites stay standard deviation.

I needed to build the forecasting model for the outflow of clients using available data (churn being the target variable).

We managed to train two models; logistic regression and random forest. The motivation behind including these statistical methods is based on the fact that logistic regression is a classical method for classification that gives good interpretability and is more or less stable. In contrast, random forest is robust and works well with small datasets due to its bagging sampling.

Final ROC-AUC scores on test data are as follows:
• For logistic regression - 0.652
• For random forest - 0.989

The top five major features for logistic regression are state of the state of residence, age, highest education acquired, use of internet services, and several complaints.

Random forest showed much better performance measures than logistic regression, and its major features make much more sense.

From a business perspective, age, unpaid balance, annual income (and other major features) seem to be valid features for churn rate.

The model served through FastAPI framework on Uvicorn server deployed on Heroku platform. It uses DVC for data and model versioning and metrics and plots versioning.

This project showcases CI/CD pipeline implementation using GitHub Actions and Heroku deployment.

CI includes PyTest tests and PEP8 code correspondence. DVC encapsulates different training components into stages.

This project is well-documented and has a model card description.