The Machine Learning Workflow

Eight steps, mostly not modelling

A typical project: (1) define the question and a success metric; (2) collect and understand data; (3) clean it and build features; (4) split into training and test sets; (5) train a simple baseline, then better models; (6) evaluate on held-out data and analyse errors; (7) deploy; (8) monitor and retrain as data changes. Most time goes into data and evaluation, not into choosing algorithms. A fair evaluation on data the model has never seen is the heart of the whole process.

The workflow as a checklist

Keep it beside every project.

1 question + metric    "predict churn within 30 days; measure recall at 20% precision"
2 data                 sources, size, label quality, time range
3 features             cleaning, encoding, scaling
4 split                train / validation / test, no leakage
5 baseline -> models   dummy model first, then real models
6 evaluate             held-out metrics + error analysis
7 deploy               batch job or API
8 monitor              drift, performance, retraining plan

Start with a baseline

A model that always predicts the average or the most common class tells you whether your real model is actually learning anything.

Quick check: Where does most of the effort in an ML project usually go?

• Buying GPUs
• Choosing a fancy algorithm
• Writing the user interface
• Data preparation and evaluation