Unsupervised learning: Clustering Ata Kaban The University of Birmingham axk

Unsupervised learning: Clustering

Ata Kaban

The University of Birmingham

http://www.cs.bham.ac.uk/~axk

The Clustering Problem

Unsupervised Learning

Data (input) ‘Interesting structure’ (output)

-Should contain essential traits

-discard unessential details

-provide a compact summary the data

-interpretable for humans

-…

Objective function that expresses our

notion of interestingness for

this data

Here is some data…

Formalising• Data points xn n=1,2,… N

• Assume K clusters

• Binary indicator variables zkn associated with each data point and cluster: 1 if xn is in cluster k and 0 otherwise

• Define a measure of cluster compactness as the total distance from the cluster mean:

• Cluster quality objective (the smaller the better):

• Two sets of parameters - the cluster mean values mk and the cluster allocation indicator variables zkn

• Minimise the above objective over each set of variables while holding one set fixed This is exactly what the K-means algorithm is doing! (can you prove it?)

– Pseudo-code of K-means algorithm:

Begin

initialize 1, 2, …,K

(randomly selected)

do classify n samples according to nearest i

recompute i

until no change in i

return 1, 2, …, K

End

Other forms of clustering

• Many times, clusters are not disjoint, but a cluster may have subclusters, in turn having sub-subclusters.

Hierarchical clustering

• Given any two samples x and x’, they will be grouped together at some level, and if they are grouped a level k, they remain grouped for all higher levels

• Hierarchical clustering tree representation called dendrogram

• The similarity values may help to determine if the grouping are natural or forced, but if they are evenly distributed no information can be gained

• Another representation is based on set, e.g., on the Venn diagrams

• Hierarchical clustering can be divided in agglomerative and divisive.

• Agglomerative (bottom up, clumping): start with n singleton cluster and form the sequence by merging clusters

• Divisive (top down, splitting): start with all of the samples in one cluster and form the sequence by successively splitting clusters

Agglomerative hierarchical clustering

• The procedure terminates when the specified number of cluster has been obtained, and returns the cluster as sets of points, rather than the mean or a representative vector for each cluster

Application to image segmentation

Application to clustering face images

Cluster centres = face prototypes

The problem of the number of clusters

• Typically, the number of clusters is known.

• When it’s not, that is a hard problem called model selection. There are several ways of proceed.

• A common approach is to repeat the clustering with K=1, K=2, K=3, etc.

What did we learn today?

• Data clustering

• K-means algorithm in detail

• How K-means can get stuck and how to take care of that

• The outline of Hierarchical clustering methods