Outliers are those clusters which have much smaller number
of strings in them than the average number of strings per cluster.

The clustering algorithm has been tweaked to write the outliers
out as soon as they are recognized because the number
of outliers for this dataset is very large and hence memory 
requirements to maintain the outlier tree is quite big.

I clustered the large set of 13000 odd unique strings to generate
the first round of clusters. This phase (call it phase 1) discards a lot of 
strings as outliers. In the second phase, I clustered the strings
which are discarded as outliers in phase 1. A good number of strings are still
thrown out as outliers in the second phase.

Files generated in the first phase are
	affils.tmp.clusters   - contains the clusters 
	affils.tmp.outliers   - contains strings thrown out as outliers.

In the second phase, affils.tmp.outliers is clustered using the variant of BIRCH.
Files generated in the second phase are	
	affils.outliers.clusters - contains clusters generated from affils.tmp.outliers
	affils.outliers.outliers - contains strings thrown out as outliers in phase 2. 

The relevant directories are 4d, 6d, 8d, each of these corresponds to a different
run. Each of them has 4 files named as above. 

Number of clusters:
Counting each outlier as a cluster, we have two different numbers for each 
run.  The numbers for each run are given below. The first (and smaller) corresponds
to the clusters generated in phase1, clusters generated in phase2, and outliers
of phase2. (Relevant files are: affils.tmp.clusters, affils.outlier.clusters, and
	affils.outliers.outliers)
The larger second number considers only the first phase, i.e, clusters and 
outliers from phase 1. (Relevant files are affils.tmp.clusters and affils.tmp.outliers)
4d : 3448	7334
6d : 4824	8844
8d : 4743	8451

Time taken varies from 3-4 hours for each phase.
For both phases together, it is around 7-8 hours
on a 200MHz pentium running solaris.