I'm working on designing the API for making a number of back-end machine learning programs available to a streaming data application. One of the many questions that comes up is the format to use for communicating learned models back to the application, which will then execute the models.

It looks like PMML is the most prominent effort to develop a standard for interchange of predictive models. A brief overview is Pechter's 2009 article in SIGKDD Explorations: What's PMML and What's New in PMML 4.0.
A blog post by Leidner Interoperability of Machine Learning Models with PMML discusses some pluses and minuses. One minus for me is the lack of support for sequence classifiers.

I would be curious to hear of any problems or great successes that people have had using PMML for communicating from a learning component to an application, and whether there's any notable competitors to it.

asked Dec 06 '10 at 13:52

Dave%20Lewis's gravatar image

Dave Lewis

One Answer:

Hi Dave,

PMML is currently supported by all the top commercial and open-source analytic tools. Please see: http://www.dmg.org/products.html

At Zementis, we have been very successful in deploying models expressed in PMML using ADAPA. As a matter of fact, people all over the world have been using ADAPA to deploy and execute models built in a variety of PMML-compliant model building environments. That's because since 2008, in addition to offering ADAPA on site, we also started offering it as a service on the Amazon Cloud. Hourly rates start at $1.

I recently described a situation which demonstrates the power of PMML in an article I wrote for IBM developerWorks entitled "What is PMML?". See: http://www.ibm.com/developerworks/industry/library/ind-PMML1/

Hope this helps answer your question. Also, feel free to post it on the PMML group in LinkedIn (with almost 1,000 members). To join, click: http://www.linkedin.com/groupRegistration?gid=2328634



answered Dec 06 '10 at 17:38

Alex%20Guazzelli's gravatar image

Alex Guazzelli

Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.