STA 9890 - Introduction to Unsupervised Learning

Unsupervised learning:

• Discovering patterns which we hope will hold true in future samples

Types of problems:

• Groups
• Regularity / Patterns
• Outlier Detection
• Prototypes/Archetypes

How to validate?

• No ‘magic’ like hold-out sets
• Stability principle
• Hopefully doesn’t matter too much (robustness)
• Success on a downstream class
• Often very specific strategies for individual problems

Math Review

• Eigendecomposition
• Multivariate Normal
• SVD