This is the source code for the HomeDepot's data science challenge on Kaggle [+]. Our solution gives the RMSE score of 0.456 and scored at top 3.5% spot in the leadboard.

Team:

• Arman Akbarian
• Ali Narimani
• Hamid Omid

As you can see the feature engineering part involves 4 parts:

• basic features: at xgb.py before merging other features
• extended features: at feat_eng_extended.py
• extended tf-idf: at feat_eng_tfidf.py (followed by model_xgb.py)
• basic tf-idf: done at the pipeline at xgb.py

Text preprocessing is done by preprocessing.py.

Misc Files Description:

• spell_corrector.py: builds spell_corr.py a python dictionary for correcting spellings in search query.
• data_trimming.py: a pre-cleaning process needed by spell_corrector.py and Synonyms.py
• data_selection.py: feature selection after extended features are built.
• nlp_utils.py: a NLP utility funciton and wrapper classes for quick prototyping
• ngrams.py: builts n-grams
• Synonyms.py: finds the synonyms of the words in search query
• project_params.py: sets global variables for the project, I/O directories etc...

In project_params.py edit the following to the correct path

 - ``input_root_dir``: the path to the original .csv files in your system

 - ``output_root_dir``: a path with few GB disk space avaiable for I/O

To do all of the following (1,2,3) you can simply use:

``make install``

To install the required packages execute:

``pip install -r requirements.txt``

The project needs the following python packages: -numpy -pandas -scikit-learn -nltk

You may also need to install nltk data:

• in python:

  >>> nltk.download()

• or via command line:

  python -m nltk.downloader all

• or makefile will take take care of everythin:

  make install

(Particularly nlp_utils uses WordNet data.)

The data, can be downloaded from the competition's homepage [+].

run make testutils to check if things work!

Note: This documentation and repo is under construction, I will hopefully clean up the repo and modify all of the scripts such that the pipeline works perfectly on a single machine