A comparison of different embedding methods on session based recommendation with graph neural networks

Predicting users next behavior based on their previous actions is one of the most valuable but also difficult task in the e-commerce and e-marketing fields. Recommendation systems that build upon session-based data try to bring a solution to that desire. The ultimate goal of this type of recommendation system is trying to make the best predictions about the succeeding item. Sequential order of the items within a session is also kept in mind in such systems. Recently proposed SR-GNN (Session Based Recommendation with Graph Neural Networks) has benefited from graph theory and proven its adequacy about being the state-of-Art session-based recommendation model. Furthermore, there are some parts exist that can improve the overall performance. The current model uses primitive embedding type which is the simplest way of representing the items, attributes, and their relationships between each other. This study brings the SR-GNN recommendation model with different types of graph embedding techniques which are widely used in a variety of research areas. Aim of this research is investigating the the effect of the embedding types to the SR-GNN. The proposed variety of embedding techniques that applied to SR-GNN show similar but slightly worse results compared to the original SR-GNN embedding. The experimental results obtained on two real datasets show that the performance of the SR-GNN model is not affected by the embedding models and the power of the model comes from the gated graph neural network model.