Supervised and unsupervised learning

Units and bites

1. Unit 1·5 bites·About 15 min

   Machine learning basics

   Supervised and unsupervised learning: labels, regression vs classification, train and test, overfitting and clustering.

   1. Supervised vs unsupervised learning

      Learning from examples that come with answers, or finding structure without them.

      3 min

   2. Regression and classification

      The two forms of supervised learning: predicting a number or predicting a label.

      3 min

   3. How a model learns: train and test

      A model learns from part of the data and is tested on a part it has never seen.

      3 min

   4. Overfitting

      The model that memorises its training data and trips on new data.

      3 min

   5. Clustering: finding groups without labels

      Unsupervised learning in practice: grouping similar customers, then naming the groups.

      3 min
2. Unit 2·5 bites·About 15 min

   Model evaluation

   Honest data splits, the confusion matrix, precision and recall, and explaining them to the business.

   1. Train–test and time splits

      When a random split is fine, and when data must be cut by time.

      3 min

   2. The confusion matrix

      Four boxes that explain every kind of right and wrong.

      3 min

   3. Precision and recall

      How often a flag is right, and how many cases get caught.

      3 min

   4. Cross-validation

      Scoring a model several times so one lucky split does not fool you.

      3 min

   5. Explaining a model to the business

      From metric numbers to decision language: how many people, how many rupiah, how much risk.

      3 min
3. Unit 3·5 bites·About 15 min

   Common supervised models

   Linear regression, logistic regression and decision trees — when to use which, and the features that make them useful.

   1. Linear regression in practice

      Reading coefficients and residuals, and when a straight line is enough.

      3 min

   2. Logistic regression

      The first classification model worth trying: probabilities, thresholds and coefficients you can explain.

      3 min

   3. Decision trees and ensembles

      One tree is easy to read; hundreds of trees are more accurate.

      3 min

   4. Features that matter

      Recency, frequency, monetary, and calendar features such as Ramadan and payday.

      3 min

   5. Choosing a first model

      Start from a simple baseline, and move up only when the test score improves.

      3 min
4. Unit 4·4 bites·About 12 min

   Clustering in practice

   K-means step by step, scaling columns, choosing k and reading the result.

   1. K-means step by step

      Place centres, assign points, move the centres, repeat until nothing changes.

      3 min

   2. Scaling before clustering

      Why a rupiah column can drown out a frequency column, and how to even them out.

      3 min

   3. Choosing k

      The elbow method, the silhouette score, and the business question that matters more than both.

      3 min

   4. Reading clusters

      From group numbers to profiles you can explain and name.

      3 min
5. Unit 5·3 bites·About 9 min

   Dimensionality reduction

   Summarising dozens of columns into a few without losing what matters.

   1. Too many columns

      Why columns that resemble each other confuse models and are hard to plot.

      3 min

   2. PCA, plainly

      Finding the directions with the most spread, and reading explained variance.

      3 min

   3. Seeing high-dimensional data

      Projecting data into two dimensions to look at it — and what the projection loses.

      3 min

Read while you wait

Guides from the Datamew team, free on the web today.