Skip to main content

Hierarchical Clustering

In our discussion of K-means clustering, we have clustered the classes in the same level. However, in real life situation, the data exists where the classes are on different levels. Maybe the overall dataset has three clusters, then one of the clusters may have two subcategories etc. This is where hierarchical clustering comes into play.

Single linkage algorithm

This algorithm works in a bottom-up fashion. We start with the bottom of the cluster tree. We consider all the data points are its own cluster. Then we merge the two nearest points, then the next two nearest (it could be a merged node in the previous step), and so on.

One important question here is: how do we define the distance function? Secondly, how do we define the distance between two clusters (once multiple points are merged)? In case of single linkage, the distance between two clusters are defined as the distance between two closest points in those two clusters:

dist(C,C)=minxC,xCxx\text{dist}(C, C') = \min\limits_{x\in C, x'\in C'} ||x - x'||

There are several flavors of linkage methods with varying definition for the distance functions.

Complete linkage

In this case, the distance between two cluster is the distance between two furthest points in those two clusters:

dist(C,C)=maxxC,xCxx\text{dist}(C, C') = \max\limits_{x\in C, x'\in C'} ||x - x'||

Average linkage

Instead of considering the distance between two closest or furthest pairs, in popular linkage methods, somewhat average between the points in the clusters is taken. There are several different ways the average can be defined.

Distance between cluster centers:

dist(C,C)=min(C)min(C)\text{dist}(C, C') = ||\min(C) - \min(C')||

Average pairwise distance between points in two clusters:

dist(C,C)=1CCxCxCxx\text{dist}(C, C') = \frac{1}{|C|\cdot|C'|}\sum\limits_{x\in C} \sum\limits_{x'\in C'}||x - x'||

Ward's method: increase k-means cost from merging the clusters:

dist(C,C)=CCC+Cmin(C)min(C)2\text{dist}(C, C') = \frac{|C|\cdot|C'|}{|C| + |C'|}||\min(C) - \min(C')||^2