Research On Session-Based Recommendation Method Based On Graph Neural Networks And Self-Supervised Learning

With the continuous development of the digital economy,the era of information overload has arrived,and the recommendation system,as an infrastructure tool that can effectively alleviate information overload,plays an increasingly important role in information consumption,service and decision-making.In recent years,recommendation systems have been widely used in web search applications,e-commerce platforms,and online entertainment platforms.On these applications,users sometimes perform brief operations without logging into the system,and these behaviors performed by anonymous users on the system are usually organized into sessions.Session-based recommendation predicts the user’s subsequent interaction behavior based on the user’s historical interaction behavior.Unlike traditional collaborative filtering or content-based recommendation systems,session-based recommendation has received widespread attention in recent years for modeling dynamic information about user preferences and product popularity over time.With the continuous development of deep learning methods such as graph neural networks and self-supervised learning,the recommendation accuracy and user satisfaction of the session-based recommendation system have been significantly improved.However,in some practical online recommendation scenarios,due to the sparsity of data,it is difficult for existing methods to capture complex and diverse item transition information.In addition,existing session-based recommendation models often rely on a single item prediction loss function,such as the crossentropy loss function,to learn model parameters or data representations.Since such models use the only parameters involved in optimizing the target learning model,they place too much emphasis on the final performance of prediction and do not capture the dynamic information of the user behavior sequence in the session.Besides,they do not comprehensively consider the correlation and fusion of context data and item transition patterns in data representation.Finally,existing session-based recommendation models are often built on the implicit feedback behavior sequence of users,which are more susceptible to noisy data because they often ignore correlations between users with similar behavior patterns.Based on the above research status,this thesis first studies a session-based recommendation model of contrastive graph self-attention network to solve the problem of session data sparsity.In addition,considering that session-based recommendation needs to excavate complex item transition information in user behavior sequences,this thesis proposes a session-based recommendation model for a contrasting multi-level graph neural networks for explicitly capturing the high-order item transition patterns between pairs within and between sessions.The main innovation work in this thesis can be summarized as follows:(1)This thesis proposes a contrastive graph self-attention network for session recommendation,which uses different levels of graph encoders and self-attention sub-networks to model user preferences,and uses contrastive learning in the constructed user preference views at different levels.Mining self-supervised signals to alleviate the data sparsity problem.Specifically,this thesis designs three different types of graph encoders to capture different levels of item transition patterns in a session,and aggregates item representations related to the current session through an attention-based fusion model to obtain a collaborative session representation.At the same time,this thesis designs a Transformer-based self-attention subnetwork model to learn the importance of items in a session,and obtains a local session representation by performing an average pooling operation on item representations in the current session.Since the above two session representations model specific session representations from a global perspective and a local perspective respectively,this thesis further introduces a contrastive learning paradigm to maximize the mutual information between the collaborative session representation and the local session representation,enriching the model’s supervision signal and further improve the recommendation performance of the model.In this thesis,extensive experiments are conducted on three widely used benchmark datasets,and the results show that the proposed contrastive graph self-attention network outperforms the baseline methods and effectively alleviates the data sparsity problem in session recommendation.(2)In this thesis,we propose a contrastive multi-level graph neural network for session recommendation,which employs different levels of graph neural networks to model complex and high-order item transition relationships in sessions.The network first applies a session-level graph convolutional neural network and a global-level graph convolutional neural network on the current session and all sessions,respectively,to capture the pairwise item transition patterns within and between sessions,and using an attention-based fusion module to learn a pairwise relation-based session representation.Meanwhile,it adopts hypergraph convolutional neural network to capture high-order item transformation information in all sessions.The network further performs an average pooling operation on the item representations learned by the hypergraph convolutional neural network to obtain a session representation based on high-order relations.Furthermore,by applying contrastive learning in different levels of constructed session representation views,the contrastive multilevel graph neural network enables the transformation of high-order item transformation information into pairwise relation-based session representations,enhancing the construction of final session representations.mold.Extensive experiments are conducted on several widely used benchmark datasets,and the results show that the proposed method adequately captures the item transition patterns at different levels in a session.