A novel similarity metric to increase the number of rating predictions in recommender systems using hashing and node-embedding

Rating predictions tells whether or not to suggest/recommend an item to the given user. Rating predictions in the Collaborative Filtering based Recommender Systems depends on the similar-users’ ratings. In general, very few users rate the items and it is difficult to find similar-users properly. This sparsity in rating data and similar-users not being identified sometimes makes it impossible to predict the user ratings on several items which are actually relevant to the user. There are several similarity metrics proposed to improve the accuracy. However, the research on improving the number of rating predictions is very limited. This article proposes a similarity metric to improve the possibility of rating predictions by considering two factors, relevancy of the items among user-pairs and the ratings given. Min-hashing and node-embedding is used to find the first factor of the similarity metric. Second factor is known by comparing the ratings of users to items, as done usually. The proposed similarity metric is compared with other similarity metrics over two benchmark datasets and the results show that it can increase the number of rating predictions to a large extent, without reducing the accuracy.