Research On Recommendation Algorithm With Graph Neural Network

The session-based recommendation predicts the next click target based on the user’s behavior sequence,but processing session data often faces data sparsity issues.The current mainstream method uses graph neural networks to aggregate high-order neighborhoods of user and project nodes to enrich feature information.However,recommendation algorithms with graph neural networks still have problems: firstly,excessive feature information can lead to over-specialization in recommendations,resulting in user frustration.Focusing on unexpected recommendation can effectively alleviate over-specialization,but its performance depends on whether the model can thoroughly learn the target feature information.Secondly,session-based recommendations typically rely on long sessions to learn user interest preferences.Although this approach avoids the prediction challenges of short session recommendations and the negative effects caused by insufficient learning in short sessions,it also overlooks user feature information contained in short sessions.Some researchers have introduced short session context sessions as supplementary information,but they cannot guarantee to find similar contexts and effectively learning feature information in short sessions.To tackle these current problems in session-based recommendation algorithms with graph neural networks,we have undertaken the following research:The unexpected interest recommendation models often suffer from the problem of insufficient feature learning.To address this problem,we propose an unexpected interest recommender system with graph neural network(UIRS-GNN).Firstly,we use a graph neural network for data preprocessing to aggregate high-order neighborhood features of the target node.Secondly,we use a personalized recurrent neural network based on attention mechanism to learn user behavior sequences.It can discover the current user’s long-term and short-term interest preferences.Then,we use the unexpected interest model to learn the user’s unexpected interest.Finally,the user’s longshort term and unexpected interests are weighted to recommend the next item.The experimental results on three universal datasets show that the recommendation performance of the proposed model is superior to the selected baseline model.Due to the scarcity of short session information in session-based recommendations,how to learn more affluent user preferences and find similar context sessions more accurately from short sessions has become an urgent problem.To address this problem,we propose a multi-feature fusion short session recommendation model(MFF-SSR).Firstly,we use a graph neural network and gated recurrent unit to learn short-session nodes’ high-order neighborhood and sequential features.Then,we apply a customized context session retrieval network to select similar sessions as supplementary information for short sessions.Finally,we use a location-aware multi-head self-attention network to explore the target and context session’s location-aware features.We weigh both to recommend the next item for the current user.This article conducts experiments on two general session datasets,and the results show that the proposed model outperforms the selected comparison models.Our research results have improved the feature learning ability of unexpected interest recommendations and short session recommendations,making recommendation algorithms more in line with human thinking patterns and more adaptable to complex data environments.This provides users with more user-friendly and precise recommendations,expanding the application of session recommendation in a wider range of fields.