There is a MinHash distribution in Mahout. I have been looking at implementing my own and I have. The suggestion that similarity between users can be determined by the least hash from user's click history ( and thus implicit 0/1 preference of an article ) seems too narrow, even if we were to use multiple hash functions and decide a probability based on number of times the min hashes match.
Any takes on whether this approach is good for generating recommendations and any good papers that suggest any empirical evidence for the same ? -- View this message in context: http://lucene.472066.n3.nabble.com/MinHash-tp3428977p3428977.html Sent from the Lucene - General mailing list archive at Nabble.com.
