Machine Learning in Excel: How Models Really Work

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Scale two customer features of Copperleaf Coffee with min-max and z-score, using statistics from the training rows only.
• Fit a line through ten weeks of ad spend and new subscribers by hand with the normal equations, and use it to predict next week.
• Train the same kind of line with gradient descent: four steps from w = 0, b = 0, watching the cost (MSE) fall.
• Score twelve subscribers with a trained logistic regression: churn probability, predicted class and the log-loss that tells you how good the probabilities are.
• Judge the churn model on 20 test customers: confusion matrix, accuracy, precision, recall and F1, and what happens when you lower the threshold.
• Classify a new Copperleaf subscriber by finding the three most similar existing customers and letting them vote.
• Run one full k-means iteration on twelve customers: distances to three centroids, cluster assignment, new centroids and the inertia.
• Do what a decision tree does at its first node: measure the Gini impurity of two candidate splits and pick the better one.