Goal-based filtering approach for recommender system

Abstract

Recommender system is a sub part of information retrieval. It decreases the content searching time, increases the user?s interest, and provides recommendations relevant to user?s goals or interests. The major drawback of recommender system is user-based cold-start problem, which has two causes: new-user zero-rated profile recommendations and average-user low-rated profile recommendations. This research proposes goal-based filtering approach consisting of two hybrid parts; first is content-based filtering with collaborative features to overcome the first cause of user-based cold-start problem. The second is collaborative filtering with k-nearest neighbor scheme features to improve the second cause of user-based cold-start problem. The famous „MovieLens? dataset is rich with its 927 entries of user?s profile data, which makes it a choice for experiments on the proposed work. The cosine similarity and euclidean distance measurements have been used to compute the personalized profile similarities between users profile preferences according to their age, gender and occupation. These similarities are helpful to predict the recommendations to the zero-rated and low-rated users without using any extra information such as ratings, likes or dislikes. The evaluation of experiments has been performed using mean precision with result of 83.44% and mean recall with result of 85.22%. The results demonstrate that percentage of user?s profile similarity measurements is probably effective for web-based system?s recommendations.