Session-based Recommendation Using Recurrent Neural Networks: A Comparative Theoretical Analysis

Recommender Systems(RSs) are widely used for providing suggestions to users on digital platforms. Recently, there has been a significant amount of interest in the class of session-based recommender systems (SBRSs) in the RS research community. In session-based recommendation scenarios, the task is to provide suggestions to the users subject to the interactions or data available in an ongoing session. In contrast to the traditional recommendation approaches (e.g. collaborative or content-based), which model long-term preferences , the SBRSs attempt to model the short-term preferences of users. Recurrent Neural Networks (RNNs) are great tools for modelling sequential data; therefore become good choices for SBRSs. This work performs a comparative analysis of the most prominent RNN-based approaches for SBRSs. Firstly, the RNN-based approaches are classified based on the characteristics of session data (such as session length, ordering of session data, etc.) that are exploited in the recommendation process. Secondly, a concise theoretical investigation of these approaches featuring their pros and cons is provided. Further, for each of these approaches, a simplified model architecture illustrating the overall functioning of the approach is also depicted. Finally, we observe that it is difficult to comprehend the state-of-the-art for RNN-based SBRS approaches.