EMUCF: Enhanced multistage user-based collaborative filtering through non-linear similarity for recommendation systems

• Published in: Expert systems with applications 2020-12, Vol.161, p.113724, Article 113724
• Main Authors: Jain, Ankush, Nagar, Surendra, Singh, Pramod Kumar, Dhar, Joydip
• Format: Article
• Language: English
• Subjects: Accuracy; Active learning; Algorithms; Bhattacharyya coefficient; Collaboration; Collaborative filtering; Density; Entropy maximization; Machine learning; Multistage rating prediction; Non-linear similarity; Ratings; Ratings & rankings; Recommendation system; Recommender systems; Similarity; Sparsity
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2020.113724

•An Enhanced Multistage User-based Collaborative Filtering (EMUCF) algorithm is proposed.•EMUCF algorithm predicts the user’s unknown rating using Bhat_sim similarity model.•A hybrid metric is proposed to elicit the dominant users and items for dense matrix.•Incremental density-based n > 2-stage EMUCF is proposed to increase prediction accuracy. The data sparsity is an acute challenge in most of the collaborative filterings (CFs) as their performance is affected by the known ratings of target users. Recently, active learning has become a prevalent and straight forward approach to cope with the data sparsity. In this approach, the newly entered users are requested to rate certain items while they signup to the underlying recommendation system. This work proposes an Enhanced Multistage User-based CF (EMUCF) algorithm, which uses the concept of active learning and predicts the unknown ratings for target users in two stages. Here, the anonymous ratings of each intermediary stage are predicted with traditional User_CF algorithm. However, the similarity models commonly used in User_CF are not adequate to compute the similarity among users. Therefore, the most recently introduced Bhattacharyya Coefficient based nonlinear similarity model Bhat_sim is used for similarity computations; it utilizes all rating pairs of items in the final estimates of users similarities. Later, an extension of simple EMUCF, the (n > 2)-stage EMUCF is proposed to increase the prediction accuracy by progressively increasing the density of the original rating matrix. The performance of simple EMUCF and its extension is evaluated on two benchmark Movielens-100K and Movielens-1M datasets. They obtain far superior results for prediction accuracy and recommendation precision compared to several prominent competing algorithms. Finally, the potential improvement in the n-stage EMUCF algorithm is assessed by establishing the connection between rating prediction accuracy and matrix density.