Research On Recommendation Model Based On Graph Neural Network For Fusing Review Text

As our life become digital,we are faced with information overload.Recommendation systems emerge as effective mainstream solutions to the information overload and important means for consumers to find the best products or services.The data in a recommendation system usually has a graph data structure and this data feature means that the recommendation system must take into account the complex graph relationships between items.With the development of deep learning,the application of graph neural networks to recommendation systems is increasingly being researched and has become an important research direction in the recommendation systems and has attracted more and more attention.Existing graph neural network recommendation models mainly rely on explicit feedback information such as users’ historical behavior and rating,while ignoring implicit feedback information from users such as review text.Therefore,in this paper,the following research on recommendation models is performed on the topic of graph neural networks in combination with review texts:(1)To address the problem that the interaction data of recent recommendation systems based on graph neural networks usually lack semantic information,and most of the recommendation systems simply stitch together different data features,resulting in insufficient interaction and thus reducing the recommendation performance,a graph neural network recommendation model fusing review texts and node features(RTN-GNNR)is proposed,and the RTN-GNNR consists of four modules.The review texts feature extraction module proposes a Bi-GRU text analysis method that combines BERT and attention mechanisms to enable the model to focus on more useful reviews.The node feature extraction module proposes a graph neural network combined with an attention mechanism for interactive node extraction,enabling the model to have better multi-source feature extraction capability.The feature fusion module proposes a tandem FM and MLP approach to achieve interactive learning between multi-source features.The prediction module inner-produces the fused higher-order features to achieve recommendations.The experiments and analysis on five datasets from Amazon showed that RTN-GNNR improved the RMSE and MSE by 1.97% to 12.53%,and more significantly on higher sparsity data.(2)To address the problem that recommendation systems based on graph neural networks usually only consider long-term preferences but not short-term preferences,resulting in poor recommendation results,a graph neural network recommendation model based on the fusion of long-and short-term preferences(LSGNN)is proposed,and the LSGNN consists of four modules.The long-term preference and item node feature extraction module proposes a graph neural network combined with an attention mechanism to extract the long-term preference and item node features from the full interaction graph.The short-term preference feature extraction module proposes a Bi-GRU method combining BERT and attention mechanism to extract shortterm preference features of users.The item text feature extraction module proposes a CNN combined with an attention mechanism to add a title representation and a description representation of the item to the interaction data.The prediction module interpolates long and short-term preference features as well as textual features of items to achieve recommendations.The experiments and analysis on five publicly available datasets from Amazon showed that LSGNN improved RMSE and MSE by 5.98% to 17.64% with some interpretability.