Introduction of alternative forms of LDA

Question

I use Latent Dirichlet Allocation with a news body from six different sources. I am interested in the evolution of the topic, its appearance and the desire to compare how the sources are the same and differ from each other over time. I know that there are a number of modified LDA algorithms, such as the Author-Topic model, themes over time, etc.

My problem is that very few of these alternative model specifications are implemented in any standard format. Some of them are available in Java, but most of them exist only as documents for conferences. What is the best way to implement some of these algorithms yourself? I am pretty good at R and jags, and can stumble upon Python when they give it enough time. I am ready to write code, but I do not know where to start, and I do not know C or Java. Can I build a model in JAGS or Python just by having formulas from the manuscript? If so, can someone point me to an example of this? Thanks.

Answer 1

My friend's answer is below, I apologize for the language.

First I wrote a Python implementation of the collapsed Gibbs sampler seen here (http://www.pnas.org/content/101/suppl.1/5228.full.pdf+html) and traced here (http: //cxwangyi.files. wordpress.com/2012/01/llt.pdf). It was slow, like balls.

Then I used Python packaging for the C implementation of this article (http://books.nips.cc/papers/files/nips19/NIPS2006_0511.pdf). It's fast as f * ck, but the results are not as great as you would see with NMF.

But the NMF implementations I saw with scitkits, and even with the recently released NIMFA library with mean compatibility, they all explode f * ck on any significant case. My new white whale is a sliced, distributed implementation of a thing. It will not be trivial.

Answer 2

In Python, do you know PyMC ? It flexibly defines both the model and the fitting algorithm.

In addition, when starting with R and JAGS, this guide exists for using JAGS in R along with the rjags package along with a set of examples .