0:00

[SOUND] In this session we are going to introduce you

an interesting extension of hierarchical clustering method,

called BIRCH: A Micro-Clustering Based Approach.

BIRCH is an abbreviation of Balance Iterative Reducing and

Clustering Using Hierarchies.

It was developed by a group of researchers in University of Wisconsin in 1996.

The general philosophy is, it incrementally constructs

a CF tree, or called a Clustering Feature tree,

which is a hierarchical data structure for multiphase clustering.

0:48

For phase 1, essentially,

it scans the database to construct an initial in-memory CF tree.

Which is a multi-level compression of the data that

tries to preserve the inherent clustering structure of the data.

1:41

The BIRCH methods scales linearly.

That means you, it finds a good clustering with a single scan, and

then improves the quality with a few additional scans.

The clustering feature we first introduce is the CF vector in BIRCH.

2:01

The BIRCH Clustering Feature essentially is suppose you get these five

points into one cluster.

okay?

Then suppose these are the five points, their positions, okay?

Then the CF vector contains three components.

One is the number of data points.

The second is a linear sum of the points in the cluster.

The third one is square sum of the N points.

Okay.

So you probably can see the, the first one, 5 means there are 5 points.

The second one acts as a linear sum of each dimension.

Okay.

The third one actually is the squared sum of each dimension.

So the Clustering Feature essentially is the summary of the statistics of

a given sub-cluster, which you can consider the number is the zeroth,

the first one is the linear, the second one is the second moments

of the sub-cluster from the statistic point of view.

That means it will register the crucial measurements of the, for

computing cluster and utilizes storage quite efficiently.

So we can look at the,

the general concepts of centroid, radius, and diameter.

Okay.

The centroid essentially is the center of the cluster, okay?

Then, suppose we have a vector of N dimensions, x sub i.

Okay.

Then, the centroid is essentially computed by the sum of

all the points in this cluster divided by the number of points in the cluster,

so that what we get is a centroid of the cluster.

Okay.

Then the radius actually is the average distance from the member

objects to the centroid.

That essentially is every one you get a difference with the centroid,

then we use the sum of their square distance.

Divide by the number of points in the cluster.

Take their square root.

Essentially it's the square root of the average distance

from any point of the cluster to its centroid.

4:19

What is diameter?

Diameter essentially is average pairwise distance within the cluster.

That means if x i and x j is within the same cluster, so essentially what we want,

we will find is there are total n times n minus 1 pairs and

we sum up all these pairwise distance then you get the square root.

That's the diameter.

4:45

Then we look at CF Tree structure in BIRCH.

The CF Tree Structure essentially is very much like a, B+-tree.

We can do incremental insertion of the new points.

That means when the new points come, okay, we can find the closest leaf entry.

Start from the root.

Okay, then we try, traverse we find the closest entry,

we can add points to the leaf entry and update the Clustering Feature, CF.

5:14

If this entry diameter is greater than the maximum diameter,

then we'll split the leaf and if it's possibly we

even will be able to split parents based on the B+-tree algorithm.

A CF tree has two parameters, one called branching factor.

That means the maximum number of children.

Another is maximum diameter of sub-clusters stored at the leaf nodes.

Then a CF tree essentially is height-balanced tree that stores

the clustering features.

The non-leaf nodes store the sums of clustering features of their children.

5:55

So we can see BIRCH is an interesting algorithm, because it, it is

an integration of agglomerative clustering with other flexible clustering methods.

The low level we do micro-clustering.

We explore the CF feature and BIRCH tree structure.

It preserves the inherent clustering structure of the data.

6:18

At the high level we do macro-clustering.

It provides sufficient flexibility for integration with other clustering methods.

So this method act, impact to many other clustering methods and

applications for large data sets.

6:36

There are some concerns.

One is the,

the BIRCH tree is still sensitive to the insertion order of the data points.

Another is, since the leaf nodes has a fixed size,

the clustering obtained may not be as natural.

And also, the clusters tend to be spherical given the radius and

diameter measure as the major parameters.