|
I have a recommender system with about 14,000 items and 400,000 users. I am trying to find the top k nearest neighbors of items to items. As of the current implementation finding all pair-wise similarities will take to long to compute (anything within 1 or 2 hrs is fine). I understand that locality sensitive hashing can find the nearest neighbors in linear time as opposed to the quadratic pair-wise approach. However, in my limited understanding of LSH, some accuracy is lost in finding the nearest neighbors. Would the applications of LSH to find similar items be appropriate for a recommender system of my dataset size? |
|
Yes, probably, as the accuracy doesn't usually degrade all that much and item similarity is only loosely correlated with recommendation quality. |
