lecture 11 Clustering

Question 1

k means

Answer 1

Convex in space
Cluster boundaries are In the middle of centers
Can’t model covariants
Only simple cluster shapes
——
Naive implementation
Fast
Not good for large dataset

Question 2

Agglomeratie clustering

Answer 2

Start with all point in their own-> 2-> 3 -> hierarchical
Dendograms
Merging criteria: complete/ average / single/ward( smallest increase)

——————
Restrict to input topology
Fase with sparse connectivity
May link to imbalanced
Give more holistic view

Question 3

Dbscan

Answer 3

Core
Allows complex cluster shapes
Can detect outliers
Two parameters to adjust
Learn arbitrary cluster shapes

Question 4

Mixture models

Answer 4

Data is mixture of small number of known distributions
Find p(x)
Guasisian
Non-convex
Parametric density
How likely a new point is

Question 5

Bayesian infinite mixture

Answer 5

Add priors on mixture coefficients and Gaussian
Can I unselect components if they don’t contribute
Possibly more robust
Replace mixture coefficients by dirichelet process
Automatically finds number of components