In a similar way, ML algorithms can be put into a Hive UDAF. I'm working on this at the moment, and it's proved quite straightforward to integrate liblinear into a UDAF. As Igor notes, by setting the number of reducers, you can set the number of parallel learners.
Robin www.baynote.com From: Igor Tatarinov <i...@decide.com<mailto:i...@decide.com>> Reply-To: "user@hive.apache.org<mailto:user@hive.apache.org>" <user@hive.apache.org<mailto:user@hive.apache.org>> Date: Thursday, January 17, 2013 1:29 PM To: "user@hive.apache.org<mailto:user@hive.apache.org>" <user@hive.apache.org<mailto:user@hive.apache.org>> Subject: Re: question about machine learning on Hive Here is how Twitter does it with Pig: http://www.umiacs.umd.edu/~jimmylin/publications/Lin_Kolcz_SIGMOD2012.pdf We use a similar approach and I think that Pig, being somewhat lower-level with better support of nested objects, is a better tool than Hive. It should be possible to do something similar with Hive but we haven't tried. The trick is to implement the learner as a serializer. Then, the number of reducers will determine how many parallel learners (bags) you can run. igor decide.com<http://decide.com> On Thu, Jan 17, 2013 at 1:23 PM, qiaoresearcher <qiaoresearc...@gmail.com<mailto:qiaoresearc...@gmail.com>> wrote: How to run machine learning algorithms (whatever ML algorithms) directly in Hive? assume the input and output already stored as Hive tables. ps: I know mahout is available there, but would prefer run machine learning algorithms directly in Hive many thanks,
