GHRS: Graph-based Hybrid Recommendation System with Application to Movie Recommendation

“RS methods are mainly categorized into Collaborative Filtering (CF), Content-Based Filtering (CBF), and hybrid recommender system based on the input data (Adomavicius and Tuzhilin, 2005). CF models (Salah et al., 2016; Polatidis and Georgiadis, 2016; Koren and Bell, 2015) aim to exploit information about the rating history of users for items to provide a personalized recommendation. In this case, if someone rated a few items, CF relies on estimating the ratings he would have given to thousands of other items by using all the other users’ ratings. On the other side, CBF uses the user item side information to estimate a new rating. For instance, user information can be age, gender, or occupation. Item information can be the movie genre(s), director(s), or the tags. CF is more applied than CBF because it only aims at the users’ ratings, while CBF requires advanced processing on items to perform well (Lops et al., 2011).

Although the CF model is preferred, it has some limitations. One of CF’s limitations is known as the cold-start problem: how to recommend an item when any rating does not exist for either the user or the item? One idea to overcome this issue is to build a hybrid model by combining CF and CBF, where side information can be utilized in the training process to compensate the lack of ratings through it. Some successful approaches extend the Probabilistic Matrix Factorization (Adams and Murray, 2010; Salakhutdinov and Mnih, 2008) to integrate side information. However, some algorithms outperform them in the general case.

Deep learning models have already been studied in a wider range of applications due to its capability in solving many complex tasks. Recently, DL has been inspiring the recommendation frameworks and brought us many performance improvements to the recommender. Deep learning can capture the non-linear user-item relationships and catches the complicated relationships within the data itself from different data sources such as visual, textual, and contextual.

In recent years, the DL-based recommendation models achieve state-of-the-art recommendation tasks, and many companies apply deep learning for enhanced quality of their recommendation (Covington et al., 2016; Okura et al., 2017).

… All of these models have shown significant improvement over traditional models. However, the existing deep learning models have not regarded the side information about the users or items, which is highly correlative to the users’ rating. Indeed, combining deep learning and side information may help us to discover a surpass solution for the considered challenges.”