Finding Structure Without Target Labels

A dataset can contain useful structure even when no target label is attached to each row.

Suppose we record two measurements from twelve fictional machines:

• average vibration;
• average temperature.

No row says healthy, failing, type A, or type B. We only have measurements.

A clustering algorithm can group rows that are similar according to those measurements. That can help with exploration, compression, anomaly review, or forming new questions.

It does not automatically reveal the true kinds of machine.

Similarity must be defined

Most clustering methods need a way to measure how close two examples are. In a two-feature plot, Euclidean distance is one common choice.

That means the representation matters. If temperature spans 0–1000 while vibration spans 0–1, temperature may dominate the distance calculation unless we transform or scale the features.

Two machines may be physically similar but appear far apart because one feature uses large numeric units. Rescaling that feature can change the neighborhoods, which can change the clusters, without changing the machines themselves.

This is a powerful reminder: the algorithm operates on the representation we give it.

Clusters are evidence for a question, not labels from nature

A useful statement is:

Under these two features, this scaling, this distance measure, and k = 3, these observations formed three groups.

A weak statement is:

The algorithm discovered three types of machine.

The second sentence invents meaning that the algorithm never tested.

Before moving on

Name one reason you might cluster data even when you do not intend to treat clusters as permanent categories.