Research On Multi-behavior Session-based Recommendation Via Graph Neural Network

Session-based recommendation,which is an important branch of recommendation system,has attracted extensive attention from industry and academia in recent years.Compared with traditional recommendation methods,session-based recommendation does not require user privacy information and historical interactions,but can provide timely and accurate recommendations for users by capturing dynamic preferences based solely on the interaction data of current session.Existing session-based recommendation works usually involve only a single type of behavior in session and cannot capture the semantics and relations from several types of behavior(such as clicking,adding-to-cart and purchasing).Therefore,it is difficult to effectively solve the problem of data sparsity.In this paper,we conduct an in-depth study on the task of multi-behavior session-based recommendation and propose a recommendation method based on the correlation analysis of multiple behaviors.The core challenge of multi-behavior session-based recommendation is cross-behavior sequence modeling,which fully captures dynamic intentions of user and transition relations of global items under different behaviors and different granularities.Therefore,we propose a method for Multi-Behavior Session-based recommendation with Graph Neural Network,which utilizes heterogeneous item graph to characterize the global multi-behavior transition relations between items.The enhanced representation of item is obtained through the graph neural network,and then embedded through the sequence model,so that it can make full use of the interaction information of auxiliary behavior to accurately capture the dynamic preferences of users and provide recommendations for target behavior.Based on the method,to cope with the sparse scenario of short sessions,this paper further considers the relations between features of items and proposes Graph Neural Network based Hybrid Model,in which explicit features and implicit features are integrated to obtain complete transition patterns.Except heterogeneous item graph,the model also constructs heterogeneous feature graph,so as to obtain the global relations among features.In addition,for cold-start scenario that lacks content information,the paper considers the combination of few-shot learning technology,and designs Memory-augmented Meta-learning Framework for Session-based target behavior Recommendation.It obtains more generalized parameters through meta-optimization framework,and utilizes fine-tuning mechanism to achieve personalized parameters for each session.The improved model includes relevant memory structures and adopts soft-clustering method to acquire knowledge from multi-behavior sessions with similar intentions.In this paper,the above models are compared with state-of-the-art models based on two public datasets,and we analyze each important module of models.The experimental results show that the proposed models have good recommendation performance,and provide some reference value for the development of related work in the future.