| Session-based Recommendation(SBR)is a task that utilizes user implicit feedback,such as purchasing or clicking behavior,to model the transitions between items within a session to predict the next item that the user may be interested in.Early SBR algorithms were mostly implemented based on Recurrent Neural Networks(RNNs),which achieved good results.However,they only consider the oneway transitions between adjacent items,ignoring the influence of other items on user preferences.To address it,researchers introduced Graph Neural Networks(GNNs)to effectively enrich the extractable item features in sessions and consider the influence of other items on user preferences.Existing SBR algorithms based on GNNs have demonstrated good recommendation performance,but there are still some issues,including: 1)focusing on extracting the user’s recent single preference in the session,which makes it difficult to mine deep-level user preferences and has a significant impact on recommendation accuracy;2)being "powerless" to solve the cold start problem;3)only modeling global and local sessions,failing to sufficiently integrate the item features of all sessions and ignoring the potential Vanishing Gradient problem that may arise during the fusion process;and 4)not fully considering the problem of imbalanced positive and negative samples during model training,making the model susceptible to interference and affecting the accuracy of the recommendation results.Regarding the above issues,this thesis focuses on the research of SBR algorithms based on Graph Neural Networks,and the main works are as follows:(1)To extract richer user preference features and alleviate the negative impact of cold start on recommendation accuracy,an Attentive Capsule Graph Neural Networks for Session-based Recommendation(ACGNN)algorithm is proposed.Session graphs are constructed based on the historical session sequences.The feature of the session graph is extracted through Gated Graph Neural Networks,generating a low-dimensional vector corresponding to each session,and the item embedding represents each session.The Dynamic Routing Capsule Network is introduced to capture the user’s multiple preferences,improve the limitation of a user’s single preference as a session representation,and aggregate high-level item embedding through an attention mechanism to enrich the user and item features in cold start,thereby improving recommendation accuracy.Comparative experiments are conducted on two real public datasets,and the results show that the proposed algorithm has significantly improved the performance compared to the baselines,and to some extent alleviated the negative impact of cold start on SBR tasks.(2)To address the Vanishing Gradient problem in the fusion of session-item features and the imbalance of positive and negative samples in model training,an Attention-enhanced Graph Neural Networks with Global Context for Session-based Recommendation(AGNN-GC)algorithm is proposed.Global session graphs and local session graphs are constructed based on all historical session sequences.Based on the constructed global session graphs,the Graph Convolutional Networks with an attention mechanism is used to learn the global-level item embedding of all sessions.Based on the constructed local session graphs,the Graph Attention Networks is used to learn the local-level item embedding of the current session.The position information of the session item is fused with the learned global-level and local-level item embeddings,and a new attention mechanism is used to enhance the item features of the current session,finely representing the current session,which helps to alleviate the Vanishing Gradient problem that may occur in the fusion process.Focal loss is introduced as the loss function to adjust the sample weights,improving the imbalance of positive and negative samples in model training.Comparative experiments are conducted on three real public datasets,and the results show that the proposed algorithm has better performance than the baselines.(3)To further enhance the ability of the algorithm to integrate high-dimensional information,we introduce a Multi-Head Attention Mechanism to fully integrate highdimensional information,and propose a Multi-Head Attentive Graph Neural Networks with Global Context for Session-based Recommendation(MAGNN-GC)algorithm.Global and local session graphs are constructed based on historical session sequences.Graph Convolutional Networks and Graph Attention Networks are used to learn globallevel and local-level item embeddings,respectively,to improve the algorithm’s expression.To improve the information fusion module,the high-dimensional information of learned global and local item embeddings is fully integrated using a Multi-Head Attention Mechanism,thereby enhancing the representation of item features in the current session.Focal loss is used as the loss function to adjust the sample weights and solve the problem of imbalance of positive and negative samples in model training.Comparative experiments are conducted on three real public datasets,and the results show that the proposed algorithm performs better than the baselines. |
