14.1 Unsupervised Learning

Unsupervised Learning

• So far we have been doing supervised learning, where have a target we’re trying to predict.
  • “How much will these homes sell for?”
  • “How long will this person spend watching this video?”
• Unsupervised learning works when we don’t have an exact target to predict, or we want to explore relationships in the data.
  • “What general types of homes are on the market right now?”
  • “What are some different segments of our customer base?”
  • “Are there distinct types of Covid-19 symptoms?”
• Clustering is one very common type of unsupervised learning.

Clustering

Goal: put observations into groups

• Those in the same group should be similar to each other
• Those in different groups should be different.

Crucial questions:

• How many groups?
• How do we define “similar” / “different”?

Artwork by @allison_horst

Many types of clustering algorithms

Source: sklearn documentation

Some differences between clustering algorithms

• Do we need to specify number of clusters?
• Do clusters have to have specific shapes?
• Does every observation have to be in a cluster (or can there be “outliers”)?
• Does every point have to be in exactly one cluster (or can there be “fuzzy” clusters)? (“hard” vs “soft” clustering)
• How fast is it? (does it scale to large datasets?)

Impact of distance metric

• What if two items are close in one dimension, but far in another?