via Speeding up K-means Clustering with Algebra and Sparse Vectors « LingPipe Blog.

This is exactly what Apache Mahout does.  We have parallelized versions of a bunch of clustering algorithms, including k-means

Going forward, there should be several new algorithms going in post 0.1, as Isabel Drost has added some code for Winnow and Perceptron implementations and I believe Karl Wettin has some work on hierarchical clustering in place.  I also believe Ted Dunning and Jeff Eastman are working on Dirichlet clustering.  Finally, Sean Owen is always rocking on Taste’s collaborative filtering capabilities, so there will no doubt be more goodness in that regard.  As for me, I’m working on integrating clustering (Carrot2 and Mahout) into Solr and will be writing a chapter on doing text clustering in Taming Text.

Tika’s interface is very similar to SAX, so it is easy to think about extraction in terms of receiving SAX events just like you do with XML, which is also nice because it is thus streaming and doesn’t require you to load a whole document into memory before dealing with it.

I’m now in the process of incorporating Tika into Solr (see SOLR-284) and I think we’ll eventually see it hooked into the Data Import Handler (DIH) in Solr, too, such that one could easily get content from a DB or a URL with the DIH, and then extract it if it is a binary object.

Ultimately, Tika is one more piece to the puzzle when it comes to dealing with content and it fits well with my personal vision (i.e. removing my Lucene PMC hat) of what Lucene is and should become.  Namely, as we move forward beyond just search (since search is a commodity these days, thanks to Lucene), it is important to have a whole suite of tools to bring to bear on the problem of dealing with structured and unstructured data.  Thus, things like Lucene, Solr, Carrot2, UIMA, Mahout, Tika, OpenNLP and other tools all should be easily usable by text tamers (riffing on my “Taming Text” theme…)  in creating intelligent applications.  As Lucene continues to develop and grow, it should become easier and easier to build things using the Lucene Stack which should spur a new wave of ideas and opportunities for those paying attention.

First off, is SOLR-651, which implements what I am calling a Term Vector Component. The basic gist of it is that Solr can now serve up term vectors from Lucene.  For those not initiated, term vectors store the term, term frequency and, optionally, position and offset information in a document-centric way in Lucene (as opposed to the inverted index storage used for searching.)  Term Vectors are often useful for doing things besides search like highlighting, machine learning, document-document similarity.  This component can provide:

1. Term
2. Term Frequency
3. Position (based on analysis)
4. Offset (character based)
5. IDF – Inverse Document Frequency

Combining all of these things, plus a couple of other features, I think, can really enable Solr to act as a more general Text server (which is what Taming Text is going to show.)  For instance, the Analysis Request Handler can act as a Document Analyzer server, and the Luke Request Handler can provide all kinds of corpus statistics.  And I haven’t even mentioned search, faceting and spell checking yet.  Nor have I mentioned the other thing I am working on:  adding search-result and document clustering to Solr.  This is taking place on SOLR-769.  The basic implementation I have now does search result clustering using the Carrot2 open source project.  After that, I plan on adding in Mahout for document based clustering.  I also know that Tom Morton, for Taming Text, has added in OpenNLP‘s Named Entity Recognition into Solr.  Some point in the near future, I’ll put up a link to that code.

Bottom line: Solr ain’t just for search anymore!

Mahout’s newest committer, Jeff Eastman, has a new blog on Mahout and Hadoop…

To learn more about k-means, see the Mahout wiki, specifically our page on k-means.

Now, we can start building more goodness in order to make Mahout world class when it comes to machine learning.  Now people actually have some real code they can check out and compile!