This part of repository holds my submission to first Kaggle competition.

Folder Structure:

• analysis: data obtained during initial analysis of input data
• preprocessing: data obtained during preprocessing. Final preprocessed data is not included and has to be calculated separately
• src: notebooks and other scripts which go through data loading up to creating a timestamped submission.

Notebooks should be run in the following order:

1. analyse.ipynb - data analysis (who would of guessed)
2. preprocess.ipynb - preprocessing and creation of data used during training
3. train.ipynb - training neural networks used in final submission. This part may run up to 8 hours or more, so be careful.
4. predict.ipynb - ensembling final submission

Some of the files use caching in order to speed up computations (e.g. calculation of Variance Inflation Factor). If data is present (in of the folders analysis or preprocessing) it is simply loaded (and eventually sorted). If you wish to turn this mechanism off (I would advise against it), just delete cached data.

Notebooks merely display results. Core is inside various utilities functions in /src/utilities. Currently those are undocumented, but should be quite readable (documentation will be updated if needed) and one should be able to follow (more or less at least).

If you need any clarification see Contact subsection.

Some of the submissions (including the best one) will be attached via Google Drive if needed (current size of over 1GB).

All seeds (especially the one in train.ipynb should be set correctly in order to replicate my results). To save some time you can check cached outputs of notebooks which display each model and it's architecture, it's training history and so on.

In order to run this code one requires third party libraries, namely:

• pytorch (easy to guess)
• pandas
• sklearn
• skorch
• statsmodels

Exact conda versions are provided inside environment.yml and using this file is advised. For more information on environment replication see here.

If you encounter any troubles contact me via e-mail: [email protected] or create an issue in this repository and I will try to respond.