DSR 29 mini-competition

1. Description

This mini competition is adapted from the Kaggle Rossmann challenge.

Goal: Predict Sales of a given Rossmann store on a given day.

Input: train.csv and store.csv file. A description for the information contained in those files can be found on Kaggle.

Output: submission.csv file from the holdout.csv file with two columns: (store) ID and Sales Prediction.

Team members

Sina Rampe

Oskar Klaja

Aloïs Villa

2. Setup

1. From your terminal, clone the repository:

With SSH: git clone [email protected]:alovg/DSR_minicomp.git

With HTTPS: git clone https://github.com/alovg/DSR_minicomp.git

2. Create environment. Compatible python version is 3.8.12.

   1. run conda create -n minicomp python=3.8.12 pip ipykernel
   2. install jupyter notebook in minicomp environment with pip3 install jupyter
   3. python -m ipykernel install --user --name minicomp --display-name “minicomp”

3. Get requirements by running pip3 install -r requirements.txt

4. Run file

   1. from the cloned repository in your terminal, run jupyter notebook
   2. navigate to xgb_group1.ipynb in browser
   3. select "minicomp" kernel
   4. Run all cells (from taskbar: go to run> run all cells)

5. You should now have a submission.csv file in the same folder.

3. Folder structure

Once cloned, the structure of the folder should look like:

DSR_minicomp

├── data

├── development_files

├── visualisations

├── .gitignore

├── README.md

├── Requirements.txt

└── xgb_group1.ipynb

The data folder should look like:

├── holdout_b29.csv

├── store.csv

└── train.csv

4. Model description

Model

The prediction was created using a Gradient Boosted Tree (with xgboost).

Data Visualisation

For the data exploration, pandas profiling was used. The exported html profiles can be found in the visualisation folder.

Features used

Existing features (used as is):

'DayOfWeek', 'SchoolHoliday', 'CompetitionDistance', 'Promo2', 'Open', 'Promo', 'Store'.

Existing features (encoded):

'Store_encoded', 'StoreType', 'Assortment', 'StateHoliday'.

Engineered features:

'Year', ‘Month’, ‘Day’, ‘WeekOfYear’, 'PromoMonth'.

Below are the features ranked in terms of importance.

Scoring

On kaggle, this model was evaluated on the root mean square percentage error (RMSPE).

RMSPE = .164

Other improvement attempts

The following are attempts which did not result in an improvement of the general performance of the model:

Models:

• Mean Lazy Estimator
• Linear Regression
• Random Forest Regression
• Tested Gradient Boosted Trees (xgBoost) parameters with validation set

Feature Engineering:

• one-hot encode all features
• target encode all features
• combination of the above two
• cyclic encoding with sinus and cosinus for DayOfWeek, day, month, year.
• upsampling store and assortment (see figures 2 and 3 below).
• imputing values for CompetitionDistance. 75km (maximum value for the feature) was set for the missing values (see figure 1 below).