An Improved Recommendation System based on Neural Matrix Factorization

This paper presents a comprehensive framework of the Neural Collaborative Filtering (NCF) approach used in the deployment of personalized recommendation systems. Collaborative Filtering (CF) techniques have resulted in the most impressive results in terms of predicting rating and providing good recommendations. However, with the advent of Artificial Intelligence and Deep Learning, the traditional CF has been replaced with NCF combined with the Matrix Factorization (MF) based models called NeuMF. Further, with the growth of the Web and its number of users and items, instead of only one criteria determining the recommendation, multi criteria recommender systems have been developed to provide more personalized recommendation. In this work, we propose a novel approach to NeuMF for multi-criteria sentiment based recommender systems which reduces computation complexity and also improves prediction accuracy. The proposed Dual NeuMF (DNeuMF) approach includes simultaneous training of two NeuMF models that eliminates the need of an aggregate function which is commonly used in multi criteria recommender systems. The entire architecture is fine-tuned through Bayesian-Lipschitz Optimization, enabling the selection of optimal hyperparameters and ultimately enhancing the model’s performance. The outcome of our model tested on Amazon Fine Foods dataset, Yelp dataset, Amazon Product dataset in terms of different evaluation parameters provides encouraging results.