I use Mahout to create a custom recommendation system that works with Boolean data.
I use GenericBooleanPrefUserBasedRecommender, NearestNUserNeighborhoodand now I am trying to determine the most appropriate user likeness function.
It has been proposed to use either LogLikelihoodSimilarityor TanimotoCoefficientSimilarity. I tried both, and in both cases I get [subjectively evaluated] meaningful results. However, the RMSE rating for the same dataset is better than LogLikehood. The “no recommendation” number is the same in both cases.
Can someone recommend which of these similarity functions is most suitable for this case?
source
share