Automatic Topic-Modelling with Latent Dirichlet Allocation

If you have the time to watch, I've also found David Blei's talk on the LDA to be very interesting and clear as an introduction

http://videolectures.net/mlss09uk_blei_tm/

David Mimno, has interesting post on the status of Mallet http://www.mimno.org/articles/malletplans/. Of note also is the Vowpal Wabbit variational Bayes https://github.com/JohnLangford/vowpal_wabbit/wiki/Latent-Di... implementation. Spark includes a parallel LDA in the 1.3 release of mllib.

I used Mahout's LDA to topic model looots of web pages. The main issue I came across was that it just fell flat when I used LDA on different sites in the same run. I'd tried a few different types of normalization and chrome filtering, but it just didn't fly. Performed best for me when it was run on a per-corpus basis.

At Kifi (https://www.kifi.com), we use LDA quite a bit for document classification, and have open sourced our Scala library: http://eng.kifi.com/reactive-lda-library/

As with any other machine learning algorithm, don't waste your time trying it on stock market, kids. Exponentially distributed randomness in, exponentially distributed randomness out.

And obligatory http://arxiv.org/abs/1208.4429

For those interested, I created a node.js LDA library that's pretty easy to use.

https://github.com/primaryobjects/lda