Article Preview
Top1. Introduction
Retaining satisfied and happy users is a big challenge for Recommender System (RS). RS can help users find things of interest that might not be expected or have seen before. For example, a movie or a restaurant that matches user preferences but is not quite popular (Suriati et al., 2017). Recommender systems are usually classified into four models (Fakhfakh et al., 2017): content-based recommender system, Collaborative Filtering (CF), hybrid recommendation, and demographic filtering.
- •
Content based recommender system: This model is based on the comparison between items and users features. The RS recommends items which are similar to the ones that the user implicitly or explicitly interacts with previously through rating or clicking. Hence, a user profile is created and used to identify new interesting and relevant items for the user (Fakhfakh et al., 2016; 2017).
- •
Collaborative filtering: This approach uses the opinions of user’s community, or the information about the past behavior, to predict the items that are interesting for a given user (Ouhbi et al., 2018). The similarity in taste of two users is calculated based on the similarity in the rating history of the users.
- •
Hybrid recommendations: This model combines two or more algorithms to create a new and better recommender system taking advantage of the strength and weakness of each one. The combination of the different techniques generates better and/or more precise recommendations.
- •
Demographic filtering: Unlike collaborative filtering, this model generates its recommendations based on user’s demographic information. The RS classifies users under a set of demographic classes representing some demographic characteristics of users known from their nationality, age, gender, location etc.
Traditional recommender systems are insufficient when dealing with high-dimensional datasets and data sparsity. During the first decade, recommender systems have been seen to be effective and efficient in improving business performance. Hence, many algorithms were proposed to solve some drawbacks in real situations like in cold start problem situation (Wei et al., 2017; Wu et al., 2017; Hdioud et al., 2017).
Traditional CF algorithms have limited learning capacities and suffer from data sparsity problem. Recently, this problem was tackled using Cross Domain Collaborate Filtering (CDCF) which is a new way to alleviate the sparsity problem in the recommender systems (Yu et al., 2018; 2019). However, these methods cannot learn an effective high-order feature interactions between users and items, and it do not take into account implicit ratings which are in some sort an implicit interest in the item no matter what the rating was.
Recently, there are many attempts to develop more sophisticated models to solve sparsity and high dimensional problems. Deep learning has witnessed a great success in many application fields such as object recognition, speech recognition, computer vision and natural language processing. There is a tendency to use deep learning to address these problems (Ouhbi et al., 2018; Cheng et al., 2016; Guo et al., 2017; Kuchaiev & Ginsburg, 2017; Zhou et al., 2018; Song et al., 2019).