Unsupervised learning extracts information, structure, or patterns from data without labeled target values. Unlike supervised learning, it works from the data itself rather than labels that guide prediction. section 1.2 section 1.7
A central application is outlier detection: finding records that differ substantially from the majority. Machine-learning approaches often examine features jointly, making subtle multivariate cases more accessible than single-value statistical checks. section 3.0 section 1.7
Unsupervised approaches may extend statistical ideas or use broader algorithmic structures:
- Estimate multidimensional distributions with kernel density estimation or multidimensional histograms. section 3.0
- Compare records using multidimensional kth-nearest-neighbor calculations. section 3.0
- Explore tree, neural-network, clustering, compression, or Markov-model methods. section 3.0
Interpretation remains essential: an outlier is not automatically a problem or an interesting discovery. When labels accumulate, semisupervised methods can score records for project-specific interest while retaining unsupervised detectors for unknown outliers. section 1.7 section 15.4
