Unsupervised learning

Most of the world's data has no label. Logs, embeddings, clickstreams, support tickets — you often need structure first: groups of similar items, a low-dimensional view, or a flag that says "this row is weird." Unsupervised learning finds that structure without a target y.

It will not replace supervised models for decisions with clear labels. It will get you features, segments, visualizations and anomaly alerts when labels are scarce. Closely related: learning vector meaning in representation learning and embeddings.

Reach for these methods when you need segments without labeled personas, a map of a high-dimensional space, alerts for rare weirdness, or cluster ids and PC scores as inputs to a later supervised model. Do not use unsupervised learning to avoid labeling when labels are cheap and the decision is clear. A mediocre supervised model with honest evaluation usually beats a beautiful clustering that nobody acts on.

Concepts in this section

Read them in order. Each one builds on the one before.

Clustering embeddings

A common pipeline: embed support tickets, normalize vectors, run k-means or HDBSCAN, label clusters by reading top terms or exemplar texts, then route tickets or discover topics.

50k support tickets → embedding → k-means with k=40 → one cluster is 90% "refund not received," another "SSO login loop." You just built a taxonomy without labeled topics. Re-embed and re-cluster quarterly as products change.

Scale features before distance-based methods. Prefer k-means++ and multiple inits. Use PCA to compress before clustering very wide data. Keep visualization algorithms (t-SNE, UMAP) out of the serving path. For anomalies, budget analyst time to review flags.

Interview angle: derive one k-means iteration on a 2D toy, explain why you standardize, say when DBSCAN beats k-means, warn about t-SNE misuse, describe isolation forests at a high level, and relate PCA to SVD.

Common mistakes

• Optimizing silhouette when nobody can act on 12 segments. Cut to the number of plays you can run.
• Feeding t-SNE coordinates into a classifier. They are a picture, not features.
• Setting isolation-forest contamination to "the true rate". It is a quota.

Key takeaways

• k-means assigns points to k centroids. Choose k with elbow, silhouette and downstream utility. Scale inputs.
• Hierarchical clustering and DBSCAN cover tree cuts and density-based shapes plus noise.
• PCA/SVD compresses variance into orthogonal components for visualization and features.
• t-SNE/UMAP are visual tools. Do not treat plot distances as data.
• Isolation forests and reconstruction errors flag anomalies. Thresholds still need human cost logic.
• Cluster embeddings for topic discovery. Judge success by decisions, not only internal scores.

Further learning

ISL chapters on PCA and clustering with labs:

An Introduction to Statistical LearningAn Introduction to Statistical Learning scikit-learn: ClusteringClustering of unlabeled data can be performed with the module sklearn.cluster. Each clustering algorithm comes in two variants: a class, that implements the fit method to learn the clusters on trai...scikit-learn scikit-learn: Decomposing signals in components (PCA)Principal component analysis (PCA): Exact PCA and probabilistic interpretation: PCA is used to decompose a multivariate dataset in a set of successive orthogonal components that explain a maximum a...scikit-learn