An Improved Recommender System for Dealing with Data Sparsity Using Autoencoders and Neural Collaborative Filtering

Abstract

With the abundance of implicit feedback, some researchers have sought to develop Recommender Systems (RS) that are entirely dependent on implicit feedback. Implicit feedback records users’ preferences by keeping track of their behaviours, such as which products they visit, where they click, which items they purchase, or how long they spend on a web page. To elicit explicit input from the users, they are prompted by the system to rate products. When compared to explicit feedback, implicit feedback cannot accurately represent user preferences. As a result, while using implicit input for RS is more difficult, it is also more practical. Collaborative Filtering (CF) that uses traditional approaches, like Matrix Factorization (MF), considers preferences of the user where the user and item latent vectors are combined linearly. These have limited learning capabilities and are plagued by data sparsity and the cold-start problem. In an effort to address these issues, several researchers have proposed integrating a deep neural network with standard CF techniques. But the research on these techniques still remain in a sparse condition. This paper proposes an improved RS for dealing with data sparsity using Autoencoders (AE) and Neural Collaborative Filtering (NCF). AEs are integrated with NCF in the proposed system, which requires inferring user and item attributes from additional data in order to anticipate user rating.