Threshold based KNN for fast and more accurate recommendations

Recommender systems attempt to predict the preference/ratings that a user would give to an item. Traditional collaborative filtering give recommendation to a user based on its similarity of ratings with the ratings of other users in the system. But they face issues such as sparsity, cold start problem, first rater problem and scalability. In the proposed framework, a user is being recommended by filtering K random users whose similarity is crossing some threshold and applying collaborative filtering only on those users. For the users/items visiting for the first time, demographic information is used. In it, demographics of users/item visiting for the first time are compared with users/item in system and discarding that user/item if a single mismatch is found. This framework has less MAE as compared to KNN or user based collaborative filtering, takes very less time to recommend as compared to above mentioned algorithms, as only K neighbors need to be considered.