Recommendation using Random Forest I am trying to write my own recommender system. I have data set of user-item rating matrix. But I do not have profile information about either items or users. I already built pure CF using cosine and Pearson. I am trying to build a recommender system using the random forest. I read some paper but those are the random forest in model-based recommender systems. Your help would be much appreciated. 
 A: Just like you calculate cosine similarity Random Forests can be used for unsupervised learning. These unsupervised Random Forests output similarity matrices based on proximity. The only caveat is that the Random Forest similarity is based on whole data, i.e. it's not pairwise calculated, you must give it the whole data for it to learn the similarity, as it's based on the number of times two given observations ended on the same leaf in the trees. In other words, this means that you'd need to recalculate the whole proximity matrix for all data every time a new observation is added.
To read about unsupervised Random Forests, I recommend this small description on Leo Breiman and Adele Cutler site (the inventors of Random Forests).
Even wikipedia includes a section on unsupervised learning with random forests.
A: RF cannot be used for memory-based recommender system.
Memory-based methods use similarity metrics(1). Similarity metrics (e.g. cosine similarity) are unsupervised and are just executing a user defined rule.
RF is not a similarity metric. It is a supervised learning model.
(1) https://en.wikipedia.org/wiki/Collaborative_filtering#Memory-based
