A Study On Graph Neural Network With Directed And Undirected Graph Fusion For Session-based Recommendation

With the rapid development of Internet and the growth of Internet information volume,recommendation system have become the basic means to help users reduce the information overload problem and quickly explore information of interest,and the session-based recommendation problem is gradually becoming one of the main research directions.Session-based recommendation problem aims to use anonymous sessions to predict current user actions.Previous approaches have modeled session sequences and extracted user feature representations for recommendation.Although good results were achieved,complex interactions and attribute relationships were ignored,and the accuracy of capturing user preferences in a session was insufficient.In order to capture accurate user features while considering the complex interaction and attribute relationships of objects in a session,inspired by graph neural networks,we propose a graph neural network session recommendation model based on directed undirected graph fusion,referred to as SR-MGF model.Porpose a neural network with a fusion of directed and undirected graphs is used for modeling to achieve feature extraction of nodes in a session.A directed graph based on interaction relations is constructed to capture the sequence features of the session and the complex interaction features between objects,and an undirected graph based on attribute relations is constructed to capture the contextual association features of the session objects.An attention mechanism that incorporates global and local information is designed to generate session-level features with different meanings to represent the current and global interests of users.The features are supplemented with macro correlation information to better target the contingent and random features of the session scenarios,and to improve the adaptability and stability of the model for the session recommendation scenarios.Finally,Multiple experiments are conducted on two real scenario datasets,and the model outperforms the mainstream comparison model in both major metrics.The model modules and parameter settings were also demonstrated to be applicable and effective through ablation and super-reference experiments.