Comparative analysis of modern algorithms for generating recommendations based on sessions, in relation to the streaming usage scenario (Streaming Session-based Recommendation)

Abstract

Recommendation systems are actively used in many areas of modern life (e-commerce, banking, communications, entertainment, etc.) and are of great importance for businesses and consumers. A separate class of these systems are session-based recommendation systems, the key feature of which is the generation of recommendations based on the user's recent actions in the system (his current session), the analysis of which allows to identify the current intentions and interests of the user. Especially relevant is the use of session-based recommendation systems in a streaming usage scenario (Streaming Session-based Recommender Systems), for example, on entertainment content platforms, marketplaces, etc. A distinctive feature of the streaming scenario is the continuous, high-volume and high-speed nature of the receipt of new data that needs to be processed in real time. In this paper, a comparative analysis of modern algorithms of session-based recommendation systems for a streaming usage scenario is carried out: Streaming Session-based Recommendation Machine, Global Attributed Graph Neural Network, Multi Global Information Assisted Streaming Session-Based Recommendation System, the general principles of building these systems, their main differences are highlighted, advantages and disadvantages are considered. Based on the research and analysis of these systems, basic (typical) recommendations for the construction of architecture, algorithms and the scenario of the work of recommendation systems based on sessions, depending on external conditions, have been developed.