RWML #023: How to solve ML problems in the real-world

3 tips to make your ML-life easier

Read time: 5 minutes

Kaggle datasets and notebooks are a great way to start in Machine Learning. There you learn things like:

• data preparation

• feature engineering

• model training

• model hyperparameter tuning.

All these steps are necessary to build good ML models. However, in a real-world ML project, there are a few extra things you need to do, that you won’t learn in Kaggle.

Here are the top 3 most recurring hidden problems I have faced in my ML life, and my tips for you to deal with them.

Tip #1. Understand the business problem first, then frame it as an ML problem 🖼️

Otherwise, you might end up building a great ML model in a Jupyter Notebook, that does NOT move the business metric it was intended to move.

This is what I call: “the perfect solution for the wrong problem”.

My advice

I always ask 3 questions at the beginning of every project:

1. What is the business outcome that management wants to improve?
   It is crucial you talk with all relevant stakeholders at the beginning of the project. They have more business context than you and can help you understand what is the target you need to shoot at.
   

2. Is there any solution currently working in production to solve this, like some rule-based heuristics?
   If there is one, this is the benchmark you have to beat in order to have a business impact. Otherwise, you can have a quick win by implementing a non-ML solution.
   

3. Is the model going to be used as a black box or as a tool to assist humans to make better decisions?
   Creating black-box solutions is easier than explainable ones. If you work in healthcare, for example, you need explainability. If you work in financial trading, you don't.

If you can answer these 3 questions it means you know WHAT is the Machine Learning problem you need to solve.

And that is a fantastic starting point for the project.

Tip #2: Focus on getting more and better data 📈

In Kaggle competitions you are given a fixed and static dataset. In fact, all participants use the same data and compete against each other on who has the better model. The focus is on models, not on the data.

In reality, the exact opposite happens.

Machine Learning models are a combination of software (e.g. from a simple logistic regression all the way to a colossal Transformer) and DATA (capital letters, yes). Data is what makes projects successful or not, not models.

But, how do you get more and better data?

My advice

These are the 2 things I recommend you do:

1. You need to talk (a lot) with the data engineering guys.
   They know where each bit of data is. They can help you fetch it and use it to generate useful features for your model. Also, they can build pipelines to add 3rd party data that can help you.
   

2. You need to be fluent in SQL.
   The most universal language to access data is SQL, so you need to be fluent in it. This is especially true if you work in a less data-evolved environment, like a startup. Knowing SQL lets you quickly build the training data for your models.

Tip #3. Structure well your code 🏗️

Jupyter notebooks are great for quickly prototyping and testing ideas. Python is a language designed for fast iterations, and Jupyter notebooks are the perfect match. However, notebooks quickly get crowded and unmanageable.

It is best practice to structure your Python code as a package and avoid code duplication.

My advice

Python Poetry is my favorite packaging tool. With just 3 commands you can generate most of the scaffolding you need.

$ poetry new my-package --name src
$ cd my-ml-package && poetry install

Poetry installs your local package as an editable Python dependency in your virtual environment, which means you can

• define Python functions inside src/my_file.py, and

• import them wherever you need the, for example, in a notebook, by running from src.my_file import my_function

At the end of the day, every ML project I work on has a structure like this:

my-ml-package
├── README.md
├── data
│   ├── test.csv
│   ├── train.csv
│   └── validation.csv
├── models
├── notebooks
│   └── my_notebook.ipynb
├── poetry.lock
├── pyproject.toml
├── queries
└── src
    ├── __init__.py
    ├── inference.py
    ├── data.py
    ├── features.py
    └── train.py

That’s all for today folks.

I hope you learned something new.

Enjoy your day!

And see you next week.

---

Whenever you're ready, there are 2 ways I can help you:

1. If you are still struggling to understand how to build, deploy and monitor a complete ML system, I’d recommend starting with a hands-on tutorial.

   
   → The Real-World ML Tutorial is a hands-on tutorial, that shows from A to Z how to build, deploy and monitor a real-world ML product. You will also join my private Discord server, so you can ask me questions and connect with other students. Join here.

2. Do you need help building an ML project or landing your first ML job?
   Book a -on-1 session with me and get a personalized action plan.
   Book your slot here.