4 ms·

Automatic Topic-Modelling with Latent Dirichlet Allocation

44 points by cschmidt 12y ago

jaydub 12y ago
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/ http://videolectures.net/mlss09uk_blei_tm/
atto 12y ago
At Kifi (https://www.kifi.com 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/ http://eng.kifi.com/reactive-lda-library/
charlescearl 12y ago
David Mimno, has interesting post on the status of Mallet http://www.mimno.org/articles/malletplans/ http://www.mimno.org/articles/malletplans/. Of note also is the Vowpal Wabbit variational Bayes https://github.com/JohnLangford/vowpal_wabbit/wiki/Latent-Dirichlet-Allocation https://github.com/JohnLangford/vowpal_wabbit/wiki/Latent-Di... implementation. Spark includes a parallel LDA in the 1.3 release of mllib.
MoOmer 12y ago
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.
dchichkov 12y ago
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 http://arxiv.org/abs/1208.4429
- upquark 12y ago
  What exactly do you mean by exponentially distributed randomness? LDA is typically applied to textual data, so the application would be to e.g. automatically tag/classify financial news. I don't think the prospects of ML are as grim in this space.
primaryobjects 12y ago
For those interested, I created a node.js LDA library that's pretty easy to use. https://github.com/primaryobjects/lda https://github.com/primaryobjects/lda

44 points by cschmidt 12y ago

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/ http://videolectures.net/mlss09uk_blei_tm/

atto 12y ago

At Kifi (https://www.kifi.com 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/ http://eng.kifi.com/reactive-lda-library/

charlescearl 12y ago

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

MoOmer 12y ago

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.

dchichkov 12y ago

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 http://arxiv.org/abs/1208.4429

upquark 12y ago
What exactly do you mean by exponentially distributed randomness? LDA is typically applied to textual data, so the application would be to e.g. automatically tag/classify financial news. I don't think the prospects of ML are as grim in this space.

primaryobjects 12y ago

For those interested, I created a node.js LDA library that's pretty easy to use. https://github.com/primaryobjects/lda https://github.com/primaryobjects/lda