Article Preview
Top1. Introduction
Nowadays, People are using the internet in their everyday tasks more than ever (Borg & Smith, 2018). Thus, users get a large volume of information every time they search for something online. Accordingly, this big amount of information overwhelmed users (Mezni & Fayala, 2018), (Kunaver & Požrl, 2017). Therefore recommender systems are important to suggest the most suitable items for users (Lacerda, 2017). The recommender systems intended to automatically understand the user’s inclinations and provide them with an item list that they will mostly like (Karimi, Jannach, & Jugovac, 2018). There are three approaches to the recommender systems: Collaborative Filtering (CF) (Nalmpantis & Tjortjis, 2017), Content Based (CB) (Frolov & Oseledets, 2017) and hybrid (Sulikowski, Zdziebko, & Turzyński, 2017) as on (Figure 1).
Figure 1. Recommender systems approaches (Batchakui, 2017)
CF goal is to give the users a customized recommendation based on “votes” made by others in form of ratings (Fu, Qu, Moges, & Lu, 2018), (Pang et al., 2018), (F. Wang et al., 2017). It is remarked by the ability to provide new recommendations to users that vary from what they have already chosen or seen before (Xie, Chen, Shang, Huang, & Li, 2015), (Valcarce, Parapar, & Barreiro, 2018), (Pang, Jin, Zhang, & Zhu, 2017). Whereas in CB systems, a user can receive recommendation without any assistance from others. It is suitable for finding different user’s interests (Champiri, Shahamiri, & Salim, 2015), (Xu, Dutta, Datta, & Ge, 2017), (Mezni & Fayala, 2018), (Jang, Yang, Kim, & Park, 2018). It relies on finding item features, and create a user profile based on her preferences (Shu, Shen, Liu, Yi, & Zhang, 2018).
Traditional recommendation systems suffer from multiple limitations. For example, to enable the CF making a good recommendation, there should be enough set of ratings, which usually does not happen (Parvin, Moradi, & Esmaeili, 2018), (Ha & Lee, 2017). This sparsity problem leads to poor recommendation (Chu & Lee, 2017). Further, most of similarity metrics use the user ratings only without taking into account any other source of information (Parvin et al., 2018). Additionally, traditional CF suffers from the cold start problem (Gonzalez Camacho & Alves-Souza, 2018). This affects the newly added users or items to the system because there is no previous information about either users or items. Therefore, it cannot find or use any past preferences. The recommendation can be poor as it will not give her any dissimilar but relevant products (Zhao, Wang, Wang, Zhou, & Jiang, 2018), (Minkov, Kahanov, & Kuflik, 2017). Additionally, CB methods also suffer from other problems such as Over-Specialization. This means the system gives a recommendation for similar items to what already rated before (Hariadi & Nurjanah, 2017).