Multi-objective item evaluation for diverse as well as novel item recommendations

Abstract

Most of the traditional Recommendation Systems (RSs) focus on recommending only the popular items as they deal with a single objective precision/popularity. However, focusing on the diversity of the items in the recommendation list is also equally important to improve its relevance to the user, i.e., it is required to view RSs as a multi-objective optimization problem. Nevertheless, owing to popularity and diversity to be conflicting with each other, it degrades the accuracy of the recommendation list. Therefore, in this work, we use a multi-objective optimization method to maintain a trade-off between the popularity and the diversity and obtain multiple trade-off solutions in a single run. We first incorporate Bhattacharyya Coefficient in an existing nonlinear similarity computation model to create a new similarity model named as Bhat_sim to increase the prediction accuracy of the exiting rating evaluation methods. Further, we formulate a multi-parent crossover mechanism NewCross in the proposed multi-objective recommendation filtering NewCrossPMOEA which preserves the order and the frequency in the parents genes to bring good objectivity in the trade-off of recommending popular and diverse items in the recommendation list. The obtained results on the Movielens dataset demonstrate that the NewCrossPMOEA performs superior in terms of average precision, diversity, and novelty to its competing methods. Moreover, the Pareto-dominance concept of NewCrossPMOEA suggests multiple recommendation solutions of diverse and novel items to the target users in a single run.