I have not checked with canopy since you don't really tell kmeans how
many to create, it's a little hidden. That's why I said I don't care
about the number, just that I'm not loosing real/important clusters.
The size of the vectors is in the data set is something like 3000 I
think. Very little pruning, just a bixo + boilerpipe crawl of a few
sites at a minimum depth. Here is the seq2sparse command I ran:
mahout seq2sparse \
-i b2/bixo-seqfiles/ \
-o b2/bixo-vectors/ \
-ow -chunk 2000 \
-x 90 \
-seq \
-n 2 \
-nv
I agree if no one else is seeing this, it may be weirdness of my own
creation.
On 5/9/12 12:24 PM, Jeff Eastman wrote:
Does this cluster reduction happen when you prime k-means with canopy?
Can you first adjust T1==T2 to get about 200 canopies and feed that to
k-means? How wide are your term vectors? Have you tried other distance
measures?
If anybody else out there is experiencing similar problems, please
chime in.
Jeff
On 5/9/12 1:07 PM, Pat Ferrel wrote:
That's what I'm doing now. Random seeds is not really the best way to
do kmeans. However my results are repeatable as far as I've gone. And
canopy wants to generate a much larger set of clusters, with a wide
range of T1 and T2 for this data set so the theory that it does not
support 30 clusters seems unlikely although the may be a fair
distance apart.
Since I've tried several times with several random seed so the "seeds
are too close" theory doesn't seem likely.
Given canopy wants to generate more clusters, the "doesn't support k
= 30" theory doesn't seem likely.
I'm not saying that there is a real problem here but when I noticed
it I had 16,000 documents and was asking for 200 clusters and got 38.
If there is some good reason for this it would be nice to find it and
report it to the user. The "good reason" might be very helpful in the
analysis. Or it could be a bug.
At least it's out there in case others are seeing lost clusters.
On 5/9/12 7:49 AM, Jeff Eastman wrote:
Paratosh is correct in his analysis. K-means can work itself into a
situation where there are some empty clusters if the initial cluster
centers are too closely spaced or if the data really doesn't support
k clusters. This is because it assigns each vector to the most
likely (closest) cluster. If two prior clusters are very close
together this can cause one of them to become empty.
Have you tried priming k-means with canopy instead of the random
sampler?
On 5/9/12 10:35 AM, Pat Ferrel wrote:
I suspect you are right Paritosh. I ran the random seed with kmean
several times on the supplied data set and always got 28 rather
than 30 clusters. I don't care so much about the number but it
might mean that some clusters are thrown out and without looking
you couldn't tell if they were important ones or not. Just upping k
to 32 doesn't really work if you still get some thrown out.
At least i think the issue is repeatable with this data.
On 5/9/12 1:14 AM, Paritosh Ranjan wrote:
Printouts of Mahout vectors prints only the non-zero elements.
So, the centers are not empty, rather they are zero.
Prima facie, I suspect that you are getting lot of empty clusters.
This might be occurring due to the combination of distance
measure, convergence threshold and distances between vectors.
Can you try to analyze and change/play around with these parameters?
I will try to look into how the Random Cluster Initialization is
working. I will log a jira if I find some issue. However, I think
that there will be no problem in cluster initialization part.
On 09-05-2012 03:21, Danfeng Li wrote:
I got the same issue. What I found is that the initial centers
have many empty ones, the final number of clusters are decided by
the number of nonempty centers.
Here are some example of my cases:
...
CL-34358205{n=0 c=[] r=[]}
CL-34358207{n=0 c=[] r=[]}
CL-34358209{n=0 c=[] r=[]}
CL-34358213{n=0 c=[0:1.000] r=[]}
CL-34358215{n=0 c=[] r=[]}
CL-34358216{n=0 c=[] r=[]}
CL-34358217{n=0 c=[] r=[]}
CL-34358220{n=0 c=[] r=[]}
CL-34358221{n=0 c=[] r=[]}
CL-34358222{n=0 c=[] r=[]}
CL-34358223{n=0 c=[] r=[]}
CL-34358224{n=0 c=[] r=[]}
CL-34358227{n=0 c=[0:1.000] r=[]}
CL-34358228{n=0 c=[] r=[]}
CL-34358229{n=0 c=[] r=[]}
...
Is it the case there is a bug in initialization?
Thanks.
Dan
-----Original Message-----
From: Pat Ferrel [mailto:[email protected]]
Sent: Tuesday, May 08, 2012 9:13 AM
To: [email protected]
Subject: Re: kmeans not returning k clusters
Here is a sample data set. In this case I asked for 30 and got 28
but in other cases the discrepancy has been greater like ask for
200 and get 38 but that was for a much larger data set.
Running on my mac laptop in a single node pseudo cluster hadoop
0.20.205, mahout 0.6
command line:
mahout kmeans \
-i b2/bixo-vectors/tfidf-vectors/ \
-c b2/bixo-kmeans-centroids \
-cl \
-o b2/bixo-kmeans-clusters \
-k 30 \
-ow \
-cd 0.01 \
-x 20 \
-dm org.apache.mahout.common.distance.TanimotoDistanceMeasure
Find the data here:
http://cloud.occamsmachete.com/apps/files_sharing/get.php?token=0b2dacddca05c0ee48cbebd05048434425b86740
BTW when I run rowsimilarity asking for 20 similar docs I get a
max of
20 but sometimes many less. Shouldn't this always return the
requested number? I'll post this question again to the the
attention of the right person.
On 5/8/12 6:15 AM, Paritosh Ranjan wrote:
I looked at the 0.6 version's code but was not able to find any
reason.
If possible, can you share the data you are trying to cluster along
with the execution parameters?
You can also open a Jira for this and provide the info there.
On 07-05-2012 19:45, Pat Ferrel wrote:
0.6
I take it this is not expected behavior? I could be doing
something
stupid. I only look in the "final" directory. Looking in the
others
with clusterdump shows the same number of clusters and I
assumed they
were iterations.
On 5/7/12 1:21 AM, Paritosh Ranjan wrote:
Which version are you using ? 0.6 or the current 0.7-snapshot?
On 07-05-2012 02:19, Pat Ferrel wrote:
What would cause kmeans to not return k clusters? As I tweak
parameters I get different numbers of clusters but it's usually
less than the k I pass in. Since I am not using canopies at
present
I would expect k to always be honored but the quality of the
clusters would depend on the convergence amount and number of
iterations allowed. No?
