To me, it looks like a two step implementation.
a) Using KMeans++ for cluster initialization.
b) Using ball KMeans for clustering
Should we introduce KMeans++ intialization algorithm in mahout first and
see how it behaves by observing user reactions?
Once we are satisfied with the quality and scalability of KMeans++
initialization, we can try to plug in the ball KMeans.
This two step process will also help in finding errors easily if there
would be any.
( I understand that putting in KMeans++ and Ball KMeans together will
provide the best performance and quality, however, I also think that it
can create chaos if something goes wrong, that's why I am proposing to
go step by step ).
If you agree, then I can help by plugging in KMeans++ first.
On 09-08-2012 22:39, Ted Dunning wrote:
It is a different algorithm altogether.
See http://www.math.uwaterloo.ca/~cswamy/papers/kmeansfnl.pdf for reference.
The basic idea underlying the algorithm is two-fold:
- pick good starting points that have high likelihood of being in the
core of each cluster
- use points in the near neighborhood of the starting points to update
the centroids
With well-clusterable data, this gives a clustering algorithm that
provably achieves high quality in a single step with high probability.
With less well-clusterable data, the situation is less clear, but a
few iterations typically produces very good results.
The initialization algorithm is similar to k-means++. See wikipedia
for more about that.
On Thu, Aug 9, 2012 at 9:06 AM, Paritosh Ranjan <[email protected]> wrote:
Ted,
Is the ball k-means related to the ball tree based implementation of knn, or
its a different algorithm altogether?
On 09-08-2012 19:54, Ted Dunning wrote:
Also, the knn package has a single pass k-means implementation that
can easily handle 20,000 clusters or more. This is done by using an
approximate nearest neighbor algorithm inside the k-means
implementation to decrease the time dependency on k to roughly log k.
See http://github.com/tdunning/knn
Any help in testing these new capabilities or plumbing them into the
standard Mahout capabilities would be very much appreciated.
On Thu, Aug 9, 2012 at 7:05 AM, Ted Dunning <[email protected]> wrote:
The upcoming knn package has a file based matrix implementation that uses
memory mapping to allow sharing a copy of a large matrix between processes
and threads.
Sent from my iPhone
On Aug 9, 2012, at 1:48 AM, Abramov Pavel <[email protected]>
wrote:
Hello,
If think Zipf's law is relevant for my data. Thanks for idea about
memory-mapping.
1) How can I "drop" extremely small/large clusters? There are 50% small
clusters with only 1 member while 1 large cluster has 50% members. Is
there a way to "split" large clusters with Kmeans?
2) Can I force Mahout not to use exact centroid but the closest point
from
my set? Any point has ~10 non-zero components while exact centroid is
very
dense (~200k).
Thanks!
Pavel
09.08.12 5:43 пользователь "Lance Norskog" <[email protected]> написал:
If Zipf's Law is relevant, locality will be much better than random.
Maybe you need a Vector implementation that is backed by memory-mapped
files?
On Wed, Aug 8, 2012 at 12:26 PM, Abramov Pavel
<[email protected]>
wrote:
Thank you Jeff, Paritosh,
Reducing cluster count from 1000 to 100 made my day. I will also try
Canopy for initial cluster count.
Unfortunately I don't know how to reduce my 200k dictionary. There are
no
unfrequent terms.
I will provide you with Hadoop config shortly. But I am pretty sure I
can't decrease number of Mappers/Reducers per node or increase memory
limits. It will affect the whole cluster.
Thank you!
Pavel
08.08.12 16:15 пользователь "Jeff Eastman"
<[email protected]>
написал:
Consider that each cluster retains 4 vectors in memory in each mapper
and reducer, and that these vectors tend to become more dense
(through
addition of multiple sparse components) as iterations proceed. With
1000
clusters and 200k terms in your dictionary this will cause the heap
space to be consumed rapidly as you have noted. Some times you can
work
around this problem by increasing your heap size on a per-job basis
or
reducing the number of mappers and reducers allowed on each node.
Also
be sure you are not launching reducers until all of your mapper tasks
have completed.
In order to provide more help to you, it would be useful to
understand
more about how your cluster is "well tuned". How many mappers &
reducers
are you launching in parallel, the heapspace limits set for tasks on
each node, etc.
For a quick test, try reducing the k to 500 or 100 to see how many
clusters you can reasonably compute with your dataset on your
cluster.
Canopy is also a good way to get a feel for a good initial k, though
it
can be hard to arrive at good T-values in some text clustering cases.
You, can also try hierarchical clustering with reduced k to stay
under
your memory limits.
On 8/8/12 5:40 AM, Paritosh Ranjan wrote:
A stacktrace of error would have helped in finding the exact error.
However, number of clusters can create Heap Space problems ( If the
vector dimension is also high ).
Either try to reduce the number of initial clusters ( In my opinion,
the best way to know about initial clusters is Canopy Clustering
https://cwiki.apache.org/confluence/display/MAHOUT/Canopy+Clustering
)
or, try to reduce the dimension of the vectors.
PS : you are also providing numClusters twice
--numClusters 1000 \ --numClusters 5 \
On 08-08-2012 10:42, Abramov Pavel wrote:
Hello,
I am trying to run KMeans example on 15 000 000 documents
(seq2sparse
output).
There are 1 000 clusters, 200 000 terms dictionary and 3-10 terms
document size (titles). seq2sparse produces 200 files 80 MB each.
My job failed with Java heap space Error. 1st iteration passes
while
2nd iteration fails. On a Map phase of buildClusters I see a lot of
warnings, but it passes. Reduce phase of buildClusters fails with
"Java Heap space".
I can not increase reducer/mapper memory in hadoop. My cluster is
tunned well.
How can I avoid this situation? My cluster has 300 Mappers and 220
Reducers running 40 servers 8-core 12 GB RAM.
Thanks in advance!
Here is KMeans parameters:
------------------------------------------------
mahout kmeans -Dmapred.reduce.tasks=200 \
-i ...tfidf-vectors/ \
-o /tmp/clustering_results_kmeans/ \
--clusters /tmp/clusters/ \
--numClusters 1000 \
--numClusters 5 \
--overwrite \
--clustering
------------------------------------------------
Pavel
--
Lance Norskog
[email protected]
