An efficient similarity measure to alleviate the cold-start problem
Gourav Jain, Tripti Mahara · 2019
The objective of developing a recommender system is to aid users by recommending products that might be of interest to them. In this, the collaborative filtering technique is one of the widely used methods where similarities are calculated among users/items to provide personalized recommendation. In order to calculate the similarity, various similarity measures are used. Most of these similarity methods do not perform satisfactorily in the presence of cold start users. A user is considered cold start if he/she has rated less than twenty items. In case of such users, the minimum available ratings data has to be utilized to recommend items. To resolve this problem, we propose a new similarity measure based on both City Block (CB) and Jaccard measure (CBJ). The co-rated items are considered by City Block while Jaccard measure considers the common items for similarity computation. Thus, when both of these measures are combined, they consider all the co-rated and common items. The main advantage of using CBJ is the reduced computational complexity involved in finding the similarity as compared to other similarity methods. To validate CBJ, we conduct experiments on Film Trust and MiniFilm data sets. The recommendation results on Film trust data set having 872 cold start users out of1508 users and MiniFilm data set having all the 55 cold start users reveal that the proposed CBJ method outperforms other existing methods.