Unsupervised learning

1 · What it is

Unsupervised learning is machine learning without an answer key. The model receives examples that have features, the details it can measure, but no label saying what the right answer is. Its task is to find structure on its own: which examples resemble each other, which details tend to go together, and which examples sit far from everything else.

The most common version is clustering, which sorts similar examples into groups. A popular method, k-means, drops a few centre points into the data, puts every example with its nearest centre, moves each centre to the middle of its group and repeats until the centres settle. The name comes from a paper by J. MacQueen, who described splitting data into k sets and suggested uses such as grouping similar items. You choose how many groups to look for. The model never learns what a group means, so a person looks at the results and names them.

Google Cloud’s BigQuery documentation gives a concrete case. It sorts London bike-hire stations into four groups using ride length, trips per day and distance from the city centre. One group turns out to be busy central stations; another is suburban stations with longer trips. No one told the model those kinds of station existed.

Clustering is not the only job. Dimensionality reduction squeezes many features into a few that keep most of the variation; principal component analysis (PCA) is the classic method, and t-SNE draws high-dimensional data as a flat map. Anomaly detection learns where ordinary data sits and flags what lies far away. Compared with supervised learning, which predicts known kinds of answer, unsupervised learning describes data, and without an answer key its results are harder to check.