We have data but no labels — we need to discover hidden structure, groups, or patterns without guidance on what to look for.
Machine learning paradigm where models find patterns, structure, or reduced representations in unlabeled data.
- Collect unlabeled data: only input features X, no output labels
- Choose task: clustering (find groups), dimensionality reduction (compress), or anomaly detection
- Define objective: minimize reconstruction error, maximize cluster separation, etc.
- Optimize: adjust model to best achieve the objective on data
- Interpret: examine discovered patterns for meaning and utility
- No labels needed: works with raw, unlabeled datasets
- Subjective: “good” patterns depend on the application and interpretation
- Exploratory: reveals what might be interesting, not what is definitively true
- Preprocessing: often used as a step before supervised learning
- Built from: Data Modeling — unsupervised learning is a modeling approach
- Related: Supervised Learning — contrast in label requirements
- Related: EDA — both are exploratory, discovery-oriented
- Related: Feature Engineering — dimensionality reduction creates new features
- No ground truth: cannot compute accuracy or error rate on unseen data
- Cluster interpretation: assigning meaning to clusters is subjective
- Initialization sensitivity: many algorithms (k-means) depend on random starts
- Curse of dimensionality: distance-based methods fail in high-dimensional spaces