Research On Recommendation Algorithm Based On Deep Features Of Review Text

At present,research based on reviews has achieved great success in both academia and industry,but there are still some problems.These existing technologies mainly extract the potential representations of users and items in an independent and static manner,without considering the special characteristics of each candidate,which results in the inability to fully capture the user's preferences.In addition,most of them use simple models to find useful feature combinations,without considering the deep-level interaction between users and items,and failing to mine high-level features,making the final recommendation effect poor.This paper conducts in-depth research on these issues,the main contents are as follows:Firstly,in order to dynamically characterize the interaction between users and items,a context-aware recommendation model based on dynamic characterization(CADR)is proposed.The model uses user's reviews and corresponding item's reviews as context,combines the attention mechanism to build a deep learning network framework,and dynamically learns the interactive attention characteristics of users and items.In order to make predictions more accurately,the hidden features of the scores are fused into the model.And for the problem of how to extract deep-level interaction features,the factorization machine with the attention mechanism is used to carry out deep-level feature interaction.Secondly,in order to explore the deep interaction characteristics between users and items in more depth,a recommendation model of extremely deep factorization machine that integrates attention(A-exFM)is proposed.From the perspective of depth,the model learns the explicit high-order and implicit high-order by deep learning,and adds attention mechanism to strengthen the expressive power of features;from the perspective of breadth,it learns the high-order and the low-order interactive features.Then,the model combine these two aspects to make the final prediction.Finally,experiments were conducted on the Amazon review data set.The results show that both the CADR model and the A-exFM model can improve the accuracy of score prediction,which proves the effectiveness and feasibility of both models.