1. Hierarchical clustering

• Avoid choosing number of clusters beforehand
• Dendrograms help visualize different clustering granularities (no need to rerun algorithm)
• Most algorithm allow user to choose any distance metric (k-means restricted us to euclidean distance)
• Can often find more complex shapes than k-means or gaussian mixture model

Divisive (top-down):

start with all data in a big cluster and recursively split(recursive k-means)

• which algorithm to recurse
• how many clusters per split
• when to split vs stop, max cluster size or max cluster radius or specified number of clusters

Agglomerative (bottom-up):

start with each data point at its own cluster, merge cluster until all points are in one big cluster (single linkage)

single linkage

• initialize each point to be its own cluster
• define distance between clusters to bb the minimum distance of C1 in cluster one and C2 in clustrer two
• merge the two closest cluster
• repeat step 3 until all points are in one cluster

Dendrogram

x axis shows data points (carefully ordered).

y axis shows distance between pairs of clusters.

Path shows all cluser to which a point belongs and the order in which clusters merge.

機器學習筆記（Washington University）- Clustering Specialization-week six