This enthusiasm was especially apparent at the June 4 presentation “Multi-modal Recommendation Algorithms” by Ted Dunning, MapR’s Chief Applications Architect. Surprisingly, a major part of this recommendation/machine-learning talk involved search, in particular, the use of Apache Solr/Lucene with Apache Mahout on the MapR distribution for Apache Hadoop.
The main thrust of the talk had to do with the advantage gained by using multiple behaviors as the source of input data for building a recommendation engine. Normally in a recommendation system, you observe behaviors similar to the one you want to drive through the use of your recommender, and then you use those behaviors as your input data to build and train your model. In contrast, Ted’s multi-modal approach has two new twists:
- Use multiple types of behavior as input to a Mahout-based recommendation model.
- Use the behavioral indicators output from the Mahout step as input for Solr-based search. The search engine here is abused to provide recommendations instead of search results.
The combination of Solr + MapR meets challenges many businesses face, and I predict that the excitement seen at Buzzwords foretells a lot of new technologies that will appear using the combination of MapR/Apache Mahout/LucidWorks Solr and Big Data.
Click here view Ted's slides.