What is Unsupervised Learning?
Unsupervised Learning is a machine learning paradigm that optimizes a structure-seeking objective on inputs without task target labels, producing representations, clusters, components, densities, or anomaly scores whose meaning requires separate validation.
Quick Facts
| Created | Concept established in 1960s with clustering algorithms |
|---|---|
| Specification | Official Specification |
How It Works
Choose the objective from the question
Clustering partitions examples according to a declared feature space and similarity measure. DBSCAN defines groups through density connectivity, Hierarchical Clustering records nested merges or splits, and Spectral Clustering partitions a similarity graph.
Dimensionality reduction answers a different question. t-SNE and UMAP optimize neighborhood representations, while Isomap and LLE impose distinct manifold assumptions. Running any method because labels are absent does not prove that its geometry corresponds to business categories.
Separate unsupervised from self-supervised learning
Both can start from data without human task labels, but self-supervised learning constructs prediction targets from the data, such as masked tokens or paired views, and trains with a supervised-style loss. Unsupervised learning is the broader historical family for discovering latent structure without an externally supplied target. Modern pipelines often combine them: pretrain a representation self-supervised, then cluster or model its density.
Validate stability, usefulness, and failure slices
Internal scores such as Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index measure geometry under the chosen representation; they do not prove that clusters are real or useful.
Compare algorithms, seeds, samples, feature sets, similarity measures, and cluster counts. If labels become available, keep them out of fitting and use Adjusted Rand Index or Normalized Mutual Information only as external partition evidence.
Key Characteristics
- Uses inputs without externally supplied target labels for the task
- Still requires an explicit objective, representation, similarity, or density assumption
- Can produce clusters, components, embeddings, densities, or anomaly scores
- May be inductive for new examples or transductive for only the fitted dataset
- Has no universal accuracy measure when no reference labels exist
- Requires stability, domain, downstream, and governance validation
Common Use Cases
- Exploring customer behavior segments before defining a targeting policy
- Grouping documents for corpus inspection and taxonomy discovery
- Compressing high-dimensional features before visualization or downstream modeling
- Estimating normal operating regions for anomaly triage
- Learning latent factors from user-item interaction matrices
Example
Loading code...Frequently Asked Questions
Does unsupervised learning have no labels and no loss?
It has no externally supplied task target for each example, but it still optimizes an objective. K-means minimizes within-cluster distance, PCA preserves variance, and an autoencoder minimizes reconstruction error. The objective encodes assumptions and does not by itself establish semantic correctness.
How is unsupervised learning different from self-supervised learning?
Self-supervised methods create prediction targets from the input itself and optimize a supervised-style task, such as predicting masked tokens. Unsupervised learning is a broader family for finding latent structure without external task labels. The terms overlap historically, so state the actual objective rather than relying on the label.
How do you evaluate unsupervised learning without ground truth?
Use several kinds of evidence: internal geometry metrics, stability across seeds and resamples, expert inspection, downstream task performance, and operational outcomes. Each answers a different question. A high silhouette score only shows separation under one representation and distance measure.
Why can the same dataset produce different valid clusters?
Clusters depend on selected features, scaling, distance, algorithm, initialization, and resolution. Customer data can be grouped by value, behavior, geography, or risk, and each partition may be useful for a different task. A cluster ID has no inherent meaning until the intended use is declared and validated.
Can unsupervised preprocessing leak test information?
Yes. A transform can ignore target labels and still leak the test distribution if it is fitted before the split. Fit scalers, PCA, clustering, vocabulary construction, and other learned transforms on training data or within each cross-validation fold, then apply the frozen transform to validation and test data.