/kaggle-liberty-challenge

Script for Kaggle's Liberty Challenge contest

Primary LanguagePythonMIT LicenseMIT

kaggle-liberty-challenge

Script for Kaggle's Liberty Challenge contest.

Instructions for generating submission files:

  1. Download the .csv-files sampleSubmission.csv, test.csv and train.csv into the data-folder of this repository.

  2. Run the script train.py

    • If the build_features-flag is manually set to False, the script reads the feats-dictionary from the features-folder and will finish faster.
    • Otherwise, the features are built up from scratch.

  3. The train.py-script will generate the submission files gbm_sub.csv and mean_sub.csv in the submissions folder.

    • gbm_sub.csvscores 0.32160 on the private leaderboard and mean_sub.csv scores 0.32464 (i.e. 4th place for both)

  4. Most of the computation time is spent training the gb-regressor.

    • On an i7-4790, training the randomized lasso feature selector takes around 10 minutes, training the L1-feature selectors take around 20 minutes, and training the gb-regressor take about 70 minutes.
    • The script uses a lot of RAM in in preprocessing.
    • Also during the training of the randomized lasso feature selector, RAM usage occasionally hits 100% on a 16 GB machine, along with periodical sharp rises in CPU temps. This RAM usage can be lowered by reducing the n_jobs-parameter of the randomized lasso routine.

  5. It should be noted that these submissions are not the ones I submitted to the contest. I moved towards more complex models which performed better on the public leaderboard but much worse on the private leaderboard.

    • I figured that this model was interesting as it did quite well on the private leaderboard while being very simple.
    • The 4th place finish can be achieved with one single gradient boosting regressor fitted on a selection of roughly 20 features.