OnlineRegressor

Streamlit Cloud app for no-code regression and classification tasks. This code was developed to be used as a demonstration tool in lectures about data science for managers who do not have a background in analytics.

All files are already pre-built. Only re-run preprocess_data.py to generate files for students to use if you find any errors - train, test and test with ground truth are available at data/processed.

Running the platform

Locally, one can run the Streamlit app from root:

streamlit run Home.py

Otherwise, you can fork and deploy this repo to Streamlit Community Cloud; this is how I am using it to make the platform available for students.

Comparing student results

After students have downloaded files using the platform, trained models, and gotten the predictions as a column Prediction on a .xlsx file, they must send it to the lecturer (I have been using email).

Save the files wherever you want on your machine. Then:

• Edit the file notebooks/compare_student_models.ipynb; specifically, the variables folder and files on the first cell.

  • folder should be the directory containing the student model outputs. My default is '../data/student_outputs/true_run/', ie. inside the data folder of this repo
  • files is a list of tuples. Each tuple contains a label for each group + the filename.xlsx of their predictions. The label is what will appear in the plots. For example,

  files = [('Grupo Azul', 'azul_preds.xlsx'), 
           ('Grupo da Maria', 'maria.xlsx')]

  is a valid way.,

• If you want, you can also change some auxiliary variables. They are:

  • target_delinq = 0.1: the maximum allowed default of the portfolio (currently 10%)
  • lucro_bom_pagador = 500: value in R$ of profit from a good payer;
  • preju_mau_pagador = 300: loss in R$ due to a payer default.

• Run the notebook. The figure comparing the groups and profits should be generated below. You can then copy it (I usually just screenshot) to a slide to show to the groups.

Necessary libraries (using Python 3.12.2)

You can check the output of the pip freeze in requirements.txt. We are currently using the following libraries in Python 3.12.2:

• scikit-learn=1.3.0
• lightgbm=4.3.0
• xgboost=2.0.3
• pandas=2.2.2
• matplotlib=3.8.4
• plotly=5.22.0
• streamlit=1.34.0
• openpyxl=3.1.2
• xlsxwriter=3.2.0
• category_encoders=2.6.3
• tqdm=4.66.2
• notebook=6.4.8