Part of the solr-recommender project is a cross-recommender based on Mahout. It uses the mapreduce version of the RecommenderJob as a template and implements an XRecommenderJob. Unfortunately the key part of the algorithm—the part handled by RowSimilarityJob—is done with a simple matrix multiply. This misses out on LLR or other ways of calculating the important cooccurrences.
If someone could implement a CrossRowSimilarityJob, perhaps based on Matrix Multiply I would integrate it and we’d have a prototype for a Mahout cross recommender that is already integrated into the Solr-recommender. IMHO there are a huge number of times when a recommender could be employed to use two actions or actions on two types of items in unique applications. Some applications just need this to be effective. The current implementation does work but is like a recommender with only Cooccurrence for RowSimilarityJob—no LLR—in other words it must be significantly less effective than it could be.
