Skip to content

Latest commit

 

History

History
88 lines (51 loc) · 3.82 KB

README.md

File metadata and controls

88 lines (51 loc) · 3.82 KB
Raw

This repository hosts a collection of scripts used for generating the model behind Code Coach, my project at Insight Data Science. This project uses an ensemble regressor trained on 20 million codeforces submissions to predict future rating changes based on past user behavior, and recommends problems on Codeforces for users to solve. Problem ratings (difficulties) were previously calculated here. More information about the models can be found at Code Coach.

While the code for training the ensemble regressors is desposited here, csv files and SQL databases were not uploaded due to size constraints. Additionally, while the data used for this project is publically available through the codeforces API, I wish to respect the privacy of codeforces users and not upload everyone's submission histories onto github.

Requirements

Note that packages marked as "not in production" means that while these models were used during the project, they are not present in the final codecoach.fun website.

  • Python 2.7
  • PostgreSQL 9.6.3
  • Python packages:
    • numpy
    • scipy
    • pandas
    • sklearn
    • XGBoost
    • Flask
    • psycopg2
    • Jupyter (for running notebooks)
    • (not in production) Keras
    • (not in production) Tensorflow
    • (not in production) Statsmodels
    • (not in production) eli5
    • (not in production) requests

Subfolder structure

  • application

    This folder contains code for the Code Coach webapp. Note that for security reasons, this is the local version of the application and not the version on AWS (which connects to an Amazon RDS instance).

  • data_pipeline

    This folder contains scripts used to call and parse data from the codeforces API. Code for scraping problem statements from both topcoder and codeforces (two different competitive programming platforms) were included as well, although this data was ultimately not used in the model.

  • exploration_and_visualization

    Jupyter notebooks showing visualizations of the dataset as well as initial exploration of dataset properties.

  • linear_regression

    Code for piece-wise linear regression on smoothed user rating changes. Not used for the final webapp.

  • neural_network

    Code for using feed-forward and recurrent neural networks to fit smoothed user rating changes. Not used for the final webapp.

  • ensemble

    Code for random forest regressors and gradient-boosted regressors on unsmoothed user rating changes. The random forest regressor achieves the largest r^2 value of 0.53. However, a gradient-boosted regressor with a r^2 of 0.41 was deployed for the webapp due to its much ligher size and faster runtime (10Mb vs. 11 Gb, around 10x faster).

  • screenshots

    Image files for this readme document.

Usage

Sample output from the codecoach.fun website is provided below.

The landing page for a random user (handle erased) is shown here, with a next predicted rating change.

Codecoach provides dashboards for the user to monitor practice progress.

Codecoach provides practice suggestions.

History

2017 06 25 Initial public commit