Enhanced fractional prediction scheme for effective matrix factorization in chaotic feedback recommender systems

Abstract

The biggest wish of the e-commerce industry is to forecast and estimate the taste and repugnance of users by utilizing user ratings for various products. Chaotic users' feedback drives the e-commerce industry to develop efficient, robust, and smart recommendation systems with the capability to deal with the inconsistencies in users' rating patterns and provide extremely related, appropriate, and timely recommendations. A few fractional order-based solutions for an effective matrix factorization procedure are suggested to boost recommender systems' performance regarding recommendations speed and precision. We further investigate a newly proposed enhanced fractional stochastic gradient descent (EFSGD) technique for accelerating recommendations speed through efficient matrix factorization. The Faa di Bruno fractional derivative used in EFSGD exploits the users' historical feedback for providing accurate predictions. As a result of the growing computational complexity of recommendation models, it will a challenge for EFSGD to extract group of prominent features and ignore the cluster of redundant users' and items' latent attributes for increasing the recommendation speed and reducing the computational complexity respectively. Therefore, (a) to resolve chaotic users' ratings patterns problem and (b) reduce the complexity by identifying and selecting groups of highly correlated latent features, an innovative elastic net regularized enhanced fractional adaptive model is developed for efficient matrix factorization. The proposed model for recommender systems outperforms baseline and state-of-the-art in terms of convergence speed, computational cost, and prediction accuracy. The substantial performance of the suggested strategy is verified through various latent factors, ratings prediction-based valuation measures, learning-rates, and fractional order values. Whereas the authenticity is verified through Movie-Lens and FilmTrust datasets.