Skip to main content

Module 1 — What we look for when there are no labels

The previous course rested entirely on a column to predict. Remove it: the model no longer has a right answer to compare against, no error to minimize in the usual sense. What is left to learn? A great deal — but the objective changes in nature, and with it the way we judge the result. This module establishes that shift, because it is what makes the rest of the course coherent.

The objective shifts: from prediction to structure

In supervised learning the question was "what is the value of yy?". In unsupervised learning it becomes: "how are these data organized?" We no longer try to reproduce a known answer, but to bring out a hidden structure: natural groups, dominant axes of variation, points that resemble nothing else.

The consequence is immediate and disconcerting at first: there is no longer a single score that settles matters. Two different partitions of the same customers can both be correct — one by purchasing behavior, the other by tenure. Unsupervised learning does not give the answer; it gives possible readings, and it is up to you to arbitrate according to the intended use.

Why this is the most common case in practice

Labeling is expensive. Annotating tens of thousands of images, or having experts qualify support tickets, takes time and money, and many projects start with none of it. Raw data, on the other hand, is abundant. Unsupervised learning is therefore often the only possible entry point — and frequently the first stage of a supervised project: you explore, you understand, you group, then you decide what to label first.

The three families of tasks

This course is built around three distinct questions, and knowing which one you are asking is the first reflex:

QuestionTaskModules
"Are there groups?"clustering2 to 5, 9
"Can I describe these data with fewer variables?"dimensionality reduction6, 7
"Which points are abnormal?"anomaly detection8

Clustering groups similar observations: segmenting a customer base, grouping documents by topic. Dimensionality reduction describes the data along fewer axes while keeping the essentials: compressing, visualizing, denoising. Anomaly detection isolates atypical observations: fraud, equipment failure, manufacturing defects.

The crux: evaluating without ground truth

This is the difficulty specific to unsupervised learning, and it deserves to be stated plainly right away. Without labels, there is no correctness to measure. We rely instead on three kinds of guardrail, all used in the modules that follow:

  • internal criteria, which measure the geometric quality of the result — compact, well-separated groups (the silhouette score of module 3);
  • stability: a result that changes completely when you slightly alter the data or the random seed is not a reliable result;
  • business usefulness, the final arbiter: are the groups found interpretable, actionable, and distinct in the eyes of those who know the domain?
Distance, and therefore scaling, decides everything

Almost every method in this course rests on a notion of distance between observations. A variable expressed in euros (0 to 100,000) will mechanically crush a variable expressed in years (0 to 80): the resulting groups will reflect the former alone. Standardizing before clustering is not optional, it is a condition of validity — the same reflex as for k-nearest neighbors in module 4 of the supervised course, but here the consequences are invisible, since no score exists to reveal them.

Summary

  • Without a target, the objective moves from "predict a value" to "reveal a structure"; there is no longer a unique answer, but readings to arbitrate according to use.
  • Unsupervised learning is the common case in practice, labeling being costly; it often serves as the first exploratory stage of a supervised project.
  • Three families of tasks: clustering, dimensionality reduction, anomaly detection.
  • Lacking ground truth, we evaluate by internal criteria, stability and business usefulness — and we standardize systematically, since distance decides everything.

Next module: k-means, the reference clustering algorithm, its principle and its blind spots.