mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: AW: Incremental clustering
Date Thu, 12 May 2011 17:14:27 GMT
If I could jump in here.

All of the clustering algorithms that we have follow a roughly EM algorithm
structure and can be viewed as
estimating mixture distributions.  For the uninitiated, a mixture
distribution is one where you can pretend
the data were generated by

1) picking a number i from 1 to n

2) picking a data sample from probability distribution i

Most of the algorithms that we use work by doing something a lot like this:

   for iteration = 1 to n
      # E step
      for each data point
            hard or soft assign data point to one of the distributions

      # M step
      for each distribution (aka cluster)
            estimate the distribution parameters from the points assigned in
the E step


Initialization of the distributions can happen before the algorithm starts
(mean-shift and canopy) or
during the E step (Dirichlet process clustering).

This set of distributions can easily be considered to be a classification
model.  The E step is
normal classification and the M step is model training.


To answer the specific question, if you assume even pretty weak stability
for your incoming data,
yes, you can use the clustering from past data to define the prior state for
clustering the current
data.  With k-means, this means that we don't have to do the canopy step.
 In general, we should
keep some idea of how much data has been processed in the past so that a few
new points can't
completely rewrite the clustering, but that is pretty easily done with
k-means and dirichlet.  With
k-means, that means that we keep the count in addition to the centroid.
 With Dirichlet, it means
that we need to have an online estimator for the probability distribution.


On Thu, May 12, 2011 at 9:13 AM, Benson Margulies <bimargulies@gmail.com>wrote:

> Jeff,
>
> Could you expand a bit on the subject of models in clustering? I
> mentally simplify this into 'clustering: unsupervised; classification:
> supervised.'
>
> Is the idea here that you are going to be presented with many
> different corpora that have some sort of overall resemblance, so that
> priors derived from the first N speed up clustering N+1?
>
> --benson
>
>
> On Thu, May 12, 2011 at 12:00 PM, Jeff Eastman <jeastman@narus.com> wrote:
> > Sure, by using your old clusters as the prior (clustersIn) for the new
> clustering, you can reduce the number of iterations required to converge.
> >
> > -----Original Message-----
> > From: David Saile [mailto:david@uni-koblenz.de]
> > Sent: Thursday, May 12, 2011 8:54 AM
> > To: user@mahout.apache.org
> > Subject: Re: AW: Incremental clustering
> >
> > Thank you very much everyone! This really helped a lot.
> >
> > Here is what I am planning to do:
> > I am going to compute an initial clustering after the first crawl.
> > Then, as sites are being added to the index I will simply classify them
> using the existing clusters.
> >
> > As I expect updates to be generally very small, I will only recompute the
> clustering after some threshold has been hit, like Grant suggested.
> > As Ted pointed out, this can be done with the old clusters as input.
> >
> > Thanks again,
> > David
> >
> >
> >
> > Am 12.05.2011 um 17:35 schrieb Ted Dunning:
> >
> >> Most of these algorithms can be done in an incremental fashion in which
> you
> >> can add batches to the previous training.
> >>
> >> On Thu, May 12, 2011 at 8:30 AM, Jeff Eastman <jeastman@narus.com>
> wrote:
> >>
> >>> Most of the clustering drivers have two methods: one to train the
> clusterer
> >>> with data to produce the cluster models; one to classify the data using
> a
> >>> given set of cluster models. Currently the CLI only allows train
> followed by
> >>> optional classify. We could pretty easily allow classify to be done
> >>> stand-alone, and this would be useful in support of Grant's approach
> below.
> >>>
> >>> Jeff
> >>>
> >>> -----Original Message-----
> >>> From: Grant Ingersoll [mailto:gsingers@apache.org]
> >>> Sent: Thursday, May 12, 2011 3:32 AM
> >>> To: user@mahout.apache.org
> >>> Subject: Re: AW: Incremental clustering
> >>>
> >>> From what I've seen, using Mahout's existing clustering methods, I
> think
> >>> most people setup some schedule whereby they cluster the whole
> collection on
> >>> a regular basis and then all docs that come in the meantime are simply
> >>> assigned to the closest cluster until the next whole collection
> iteration is
> >>> completed.  There are, of course, other variants one could do, such as
> kick
> >>> off the whole clustering when some threshold of number of docs is
> reached.
> >>>
> >>> There are other clustering methods, as Benson alluded to, that may
> better
> >>> support incremental approaches.
> >>>
> >>> On May 12, 2011, at 4:53 AM, David Saile wrote:
> >>>
> >>>> I am still stuck at this problem.
> >>>>
> >>>> Can anyone give me a heads-up on how existing systems handle this?
> >>>> If a collection of documents is modified, is the clustering recomputed
> >>> from scratch each time?
> >>>> Or is there in fact any incremental way to handle an evolving set of
> >>> documents?
> >>>>
> >>>> I would really appreciate any hint!
> >>>>
> >>>> Thanks,
> >>>> David
> >>>>
> >>>>
> >>>> Am 09.05.2011 um 12:45 schrieb Ulrich Poppendieck:
> >>>>
> >>>>> Not an answer, but a follow-up question:
> >>>>> I would be interested in the very same thing, but with the
> possibility
> >>> to assign new sites to existing clusters OR to new ones.
> >>>>>
> >>>>> Thanks in advance,
> >>>>> Ulrich
> >>>>>
> >>>>> -----Urspr√ľngliche Nachricht-----
> >>>>> Von: David Saile [mailto:david@uni-koblenz.de]
> >>>>> Gesendet: Montag, 9. Mai 2011 11:53
> >>>>> An: user@mahout.apache.org
> >>>>> Betreff: Incremental clustering
> >>>>>
> >>>>> Hi list,
> >>>>>
> >>>>> I am completely new to Mahout, so please forgive me if the answer
to
> my
> >>> question is too obvious.
> >>>>>
> >>>>> For a case study, I am working on a simple incremental web crawler
> (much
> >>> like Nutch) and I want to include a very simple indexing step that
> >>> incorporates clustering of documents.
> >>>>>
> >>>>> I was hoping to use some kind of incremental clustering algorithm,
in
> >>> order to make use of the incremental way the crawler is supposed to
> work
> >>> (i.e. continuously adding and updating websites).
> >>>>>
> >>>>> Is there some way to achieve the following:
> >>>>>     1) initial clustering of the first web-crawl
> >>>>>     2) assigning new sites to existing clusters
> >>>>>     3) possibly moving modified sites between clusters
> >>>>>
> >>>>> I would really appreciate any help!
> >>>>>
> >>>>> Thanks,
> >>>>> David
> >>>>
> >>>
> >>> --------------------------
> >>> Grant Ingersoll
> >>> http://www.lucidimagination.com/
> >>>
> >>> Search the Lucene ecosystem docs using Solr/Lucene:
> >>> http://www.lucidimagination.com/search
> >>>
> >>>
> >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message