## Definition
**Unsupervised learning** discovers structure in unlabelled data $\{x_i\}_{i=1}^n$. There is no target $y$ to predict — the model finds patterns, groupings, or representations from the data alone.
## Main Sub-Tasks
### Clustering
Group similar examples. Algorithms: [[K-Means Clustering]], [[Hierarchical Clustering]], [[DBSCAN]], [[Gaussian Mixture Model]].
### Dimensionality Reduction
Project high-dimensional data into a low-dimensional space preserving structure. See [[Principal Component Analysis]], [[t-SNE]], [[UMAP]].
### Density Estimation
Learn the probability distribution $p(x)$ that generated the data. Used for anomaly detection, generative modelling, sampling.
### Association Rule Learning
Find frequent patterns and rules among items. See [[Apriori Algorithm]]. Classic example: market-basket analysis.
### Anomaly Detection
Identify examples that don't fit the learnt structure — outliers, fraud, defects.
## Evaluation Challenge
Without labels, there's no straightforward analogue of accuracy or error. Evaluation depends on the task:
- **Clustering:** silhouette score, Davies-Bouldin index, mutual information against external labels (if available).
- **Dimensionality reduction:** reconstruction error, downstream task performance.
- **Density estimation:** held-out log-likelihood.
In practice, unsupervised learning is often evaluated by its *utility for a downstream task* — does the clustering improve a customer-segmentation business metric? Does the lower-dimensional embedding speed up a search system?
## When Unsupervised Wins
- **Exploratory data analysis.** What groups exist in the data?
- **Data preprocessing.** Reduce dimensionality before a supervised model.
- **Anomaly detection** when anomalies are too rare to label.
- **Pre-training** for downstream supervised tasks (the bridge to self-supervised learning).
## The Self-Supervised Bridge
Modern foundation models — including LLMs — are trained with *self-supervised* objectives derived from unlabelled data: next-token prediction, masked language modelling, contrastive learning. The distinction between "unsupervised" and "self-supervised" is technical; the practical effect is the same: usable models without manual labels. See [[Self-Supervised Learning]].
## Related
- [[K-Means Clustering]]
- [[Principal Component Analysis]]
- [[Gaussian Mixture Model]]
- [[Self-Supervised Learning]]
- [[Supervised Learning]]