Scikit-learn is a library of machine learning algorithms. PredicitionIO does the stuff around machine learning you need to get a system into production. It provides a REST API to interact with the system and it handles data storage and retrieval as well. I think it has fewer algorithms than scikit-learn. This may or may not be a problem depending on your use case.

Since the authors claim that it is "Built on top of scalable frameworks such as Hadoop and Cascading" I guess that it is aiming to be production-ready while scikit is mostly for prototyping.

In that case then what is the added value compared to Mahout, the Hadoop-based ML framework?

I'm always extremely skeptical of such initiatives, because ML is not a magical black box where you put your data in one end and you get results on the other end. Automating the trivial parts of ML, ie. providing an API to a ML library, is a week-end project, but that in itself is useless. If you don't automate the hard parts of ML, such as feature engineering, then you're not providing any value at all.

What other frameworks are out there (other than the one you mentioned) that are worth looking into?

There's waffles http://waffles.sourceforge.net/

and vowpal wabbit https://github.com/JohnLangford/vowpal_wabbit/wiki

I think the juggernaut here is WEKA [0], I've seen it used in the real world far more than anything else.

[0] - http://www.cs.waikato.ac.nz/ml/weka/

LensKit: http://lenskit.grouplens.org which also prominently features in the currently running 'Intro to Recommender Systems' MOOC on Coursera.

The MOOC course is very interesting. Do you know if there's a way to get enrolled? It looks like the registration is closed and there's no upcoming offerings available currently. Thanks.