Shan Dou
MLND capstone project
July 2018

Link to proposal review: https://review.udacity.com/#!/reviews/1314525

Kaggle competion "TalkingData AdTracking Fraud Detection Challenge".

conda env create -f environment.yml

This operation will create a conda environment named mlnd_clean. If you wish to use a different name, please open the requirement file environment.yml and change the first line name: mlnd_clean into your preferred name.

Once all the dependencies are installed, please run the following command in your shell terminal to activate the environment

source activate mlnd_clean

To deactivate, type

source deactivate

• The following modules are installed with conda

1. numpy
2. pandas 
3. seaborn
4. sklearn
5. xgboost
6. lightgbm
7. imblearn
8. notebook

• Module for stack ensemble is installed with pip:

9. mlens

For more information about mlens, please visit its webiste.

1. Jupyter notebooks:

   • MLNDcapstone_shandou_main.ipynb: Main workbook
   • MLNDcapstone_shandou_robustness.ipynb: Companion workbook for models' robustness texts

2. Python models in ./customlib/:

   • ./customlib/preprocessing.py: data processing
   • ./customlib/modeling.py: modeling
   • ./customlib/utils.py: miscellaneous tasks such as visualization and generating result summary tables

3. Dataset:
   The raw data train.csv can be directly download from Kaggle. Out of file size concerns, only downsized training data and the original testing data are included in this repo.

   • train_sample.csv: 0.1% of the raw click records
   • train_sample_2.csv: 0.2% of the raw click records
   • test.csv: First 10 lines of the orignal test data downloaded from Kaggle. NOTE that test.csv provided by Kaggle is only used for checking data fields. In the actual implementation, testing data is instead a portion of train_sample.csv or train_sample_2.csv

4. Proposals and reports:

   • proposal.pdf: Proposal of the capstone project
   • proposal_review.pdf: Comments from proposal review
   • report.pdf: Report of the capstone project

5. Others:

   • folder images/: contains all the images used in the report
   • matplotlib style sheet stylelib/custom.mplstyle: dataviz styler used throughout this project