The ML workflow end to end

Why a workflow?

Beginners often jump straight to "which algorithm is best?". In real projects the algorithm is a small part of the work. Most failures come from a vague problem, a wrong metric, leaked data or a model nobody monitors. A fixed workflow protects you from these.

The steps

1. Define the problem: who will use the prediction, and what decision changes because of it?
2. Choose a metric that matches the cost of mistakes, for example mean absolute error in minutes, or recall for fraud.
3. Collect and clean data (module 4), and check that every feature exists at prediction time.
4. Split the data into training and test sets before you look too closely (module 6 goes deeper).
5. Build a baseline: the simplest possible prediction, such as "always predict the average" or "always predict the most common class".
6. Train one or more models on the training set.
7. Evaluate on the test set. Compare with the baseline. If you cannot beat it, stop and rethink.
8. Improve: better features, more or cleaner data, tuned settings.
9. Deploy: put the model where it makes predictions for real users or reports.
10. Monitor: watch accuracy and input data over time, and retrain when the world changes.

Baselines: your first model

A baseline is a deliberately simple prediction that sets the bar. For regression, predict the mean (or median) of the training labels for everyone. For classification, predict the majority class. If 95% of loans are repaid, a "model" that always says "repaid" is 95% accurate. A real model that reaches 95.5% is barely better, which you would never notice without the baseline.

The scikit-learn interface

scikit-learn (imported as sklearn) is the standard Python library for classic ML. Every model, called an estimator, has the same methods: fit(X, y) learns from training data, predict(X) makes predictions, and score(X, y) gives a default metric (R² for regressors, accuracy for classifiers).

from sklearn.dummy import DummyRegressor
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_absolute_error
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

baseline = DummyRegressor(strategy="mean").fit(X_train, y_train)
model = LinearRegression().fit(X_train, y_train)

print(mean_absolute_error(y_test, baseline.predict(X_test)))   # e.g. 9.8 minutes
print(mean_absolute_error(y_test, model.predict(X_test)))      # e.g. 4.1 minutes

Common mistakes

• No baseline, so a weak model looks impressive. Fix: always report the baseline next to the model.
• Tuning on the test set until it looks good. Fix: keep the test set for the final check only.
• Deploying and forgetting. Data changes (new cities, festivals, prices). Fix: monitor and retrain.
• A metric that does not match the business. 99% accuracy on fraud can still miss every fraud case.