Using novelty score of unseen items to handle popularity bias in recommender systems

Recommender systems assist users to narrow down the choices from the plethora of options available to them, by recommending options most suitable to their profile. Traditional Collaborative Filtering (CF) technique suffers from the problem of popularity bias as a result of which the items recommended lack novelty in them. The need today is, to incorporate novel items in the recommended list of items, as popular items are apparent and lack novelty. This paper proposes a new Novelty Score (NS) metric based on item frequency and inverse user frequency for unseen items. The paper also proposes Modified Collaborative Filtering Approach for Novel Recommendations (MCFNR)which utilizes this metric to generate novel recommendations. The target user is categorized as occasional or persistent in this work. Collaborative Filtering is used to generate recommendations for occasional users whereas MCFNR is used to generate recommendations for persistent users as popularity bias affect such users the most MCFNR eradicates the problem of popularity bias while keeping the relevance by introducing the concept of novelty in the recommendations. The prototype of Recommender System is developed and tested with Books data set as a case study. Results are evaluated for CF and MCFNR for persistent users using precision, recall and F-measure.