milosz-l / delivery-time-prediction

ML project - package delivery time prediction

Delivery time prediction project

Project description

The goal of the project was to build an ML model that predicts the delivery time of a product purchased at given time.

The project included the following steps:

1. Determining business objectives with the client
2. Data cleaning, EDA, data visualization, feature selection, feature engineering, etc.
3. Preparing model for deployment as a web service

Technical aspects

• Code is organized in a readable structure
• Poetry allows for easy and effective management of dependencies
• Makefile contains short commands for repeatable tasks
• Only modified components of the pipeline will be rerun
  • e.g. make pipeline won't run process data if it's not necessary - it will go straight to training
• Thanks to pre-commit, before every commit the following steps are made automatically:
  1. Code formatting using black
  2. Code quality is being checked with flake8
  3. Imports are being sorted with isort
  4. interrogate checks if there are any missing docstrings (if some must-have conditions are not met, the commit will fail - this forces to write good quality code)
• pdoc automatically generates documentation
• Prefect allows you to observe and orchestrate the data processing / training pipeline, for example:

Project structure explained

.
├── data            
│   ├── final                       # predictions
│   ├── processed                   # data after processing
│   ├── raw                         # raw data
├── docs                            # documentation
├── .flake8                         # configuration for flake8
├── .gitignore                      # define which files should be ignored by git
├── Makefile                        # store useful commands to set up the environment
├── models                          # store models
├── notebooks                       # store notebooks
├── .pre-commit-config.yaml         # configurations for pre-commit
├── pyproject.toml                  # dependencies for poetry
├── README.md                       # project description you are reading right now
├── src                             # store source code
│   ├── __init__.py                 # make src a Python module
│   ├── config.py                   # store configs 
│   ├── process.py                  # process data before training model
│   ├── run_notebook.py             # run notebook
│   └── train_model.py              # train model
└── tests                           # store tests
    ├── __init__.py                 # make tests a Python module 
    ├── test_process.py             # test functions for process.py
    └── test_train_model.py         # test functions for train_model.py

Business problem

It seems that our courier companies sometimes fail to handle deliveries. If we knew how long such delivery for a given order would take - we could pass this information on to customers.

So we want to be able to predict the delivery time of a product purchased at a given time.

Business success criteria

Expectations: The model will estimate the time of delivery and predict its date and time with an accuracy of +- 24 hours in 90% of cases.

Results: In the end, the model is correct with an accuracy of +- 24 hours in 93% of cases, so the success criteria was met.

Selected algorithm

After comparing several regression models, RandomForestRegressor was selected. The results given by this model were far better than results given by linear regression models.

Quick Start

Set up the environment

1. Install Poetry
2. Set up the environment:

make setup
make activate

Install new packages

To install new PyPI packages, run:

poetry add <package-name>

Run Python scripts

To run the Python scripts to process data and train model, type the following:

make pipeline

Web service with predictions

The web service is a Flask app that runs on port specified in src/config.py file.

To start this web service, type the following:

make app

You don't need to make pipeline beforehand. It will be run automatically only when it's needed.

Now in order to make predictions you just need to send a POST request with data as a JSON body, for example: More example requests are in the src/example_post_requests folder.

View all flow runs with Prefect

A flow is the basis of all Prefect workflows.

To view the flow runs from a UI, sign in to your Prefect Cloud account or spin up a Prefect Orion server on your local machine:

prefect orion start

Open the URL http://127.0.0.1:4200/, and you should see the Prefect UI.

Auto-generate API documentation

To auto-generate API document, run:

make docs_save

To view the docs, run:

make docs_view

Cookie Cutter template

The initial project structure has been based on this Cookiecutter template: https://github.com/khuyentran1401/data-science-template