Research On Deep Recommendation Algorithm Based On Review Text

In order to alleviate the information overload problem caused by the exponential growth of Internet information,recommendation systems are widely used in many fields.The review text contains multiple information of users and items,and the recommendation system can mine the information in the review text to predict the user's preference for the item.Recommendation systems have carried out a wealth of research on this.However,how to improve the semantic understanding of review text,extract deep-level features of review text,and capture more comprehensive user and item features deserves further research.Based on the above problems,this thesis explores the research of recommendation algorithm based on review text.Existing methods usually interact the features of users and items equivalently,but user features represent preferences for target items,and item features represent the attribute information of items,that is,the importance of users and products is different and asymmetric.Therefore,this thesis proposes an review text based adaptive feature extraction recommendation model.The model first uses the dynamic word embedding pre-training model BERT to solve the problem of word ambiguity and avoid the bias caused by semantic understanding.Secondly,the Bi-GRU network and attention mechanism are used to extract user features and item features to enhance the feature expression ability.Finally,an adaptive feature splicing mechanism is designed to balance the importance of users and items in feature interaction.Experiments on six Amazon datasets show that the model outperforms the baseline models,and the adaptive feature splicing mechanism can effectively balance the respective importance of user and item features,and improve the accuracy of prediction rating.In the past,recommendation systems based on review text usually mine static review information and model user and item features,ignoring the problem that user preferences and item popularity will shift over time.This thesis proposes an review text based long-term and Short-term Feature Extraction Recommendation.The model firstly mines the long-term features of reviews by using word-level and review-level two layer attention mechanisms.Then a temporal attention layer is designed to combine time sequence information and review text to capture the short-term features of users and items.Finally,covariance is introduced to remove redundant information between long-term features and short-term features.Experiments on four Amazon datasets show that the model outperforms the baseline models,and the long-term and short-term feature extraction network can effectively extract dynamic features of user preferences and item popularity,and improve recommendation performance.