Clustering

• Find subgroups among observations.
• Get insights or run predictions.

Clustering is used to find groups to target adds. Also used in genomic research.

K means clustering

• Centroid: average position of a cluster
• Distortion: the mean of square distance
• Inertia: the sum of the square distance

K is a hyper-parameter - the number of clusters. Plot elbow plot to find what value of k reduces distortion the most (but use lowest possible).

Vulnerable to curse of dimensionality. PCA preprocessing helps.

Given enough time, K-means will always converge (centroids stop moving per iteration)

Hierarchical clustering

• Does not need to know K in advance
• Use (di)similarity or distance metric
• Use Dendrogram
• Deterministic
• Greedy

Choice of linkage type influences how clusters are formed.

• complete
• single
• average
• centroid
• ward

Measure distance between clusters and merge if shorter than a threshold.