Research On Recommendation Algorithm Based On Comment Text

The massive information problems brought about by the Internet can easily cause users to get lost in the information and cannot find what they want.The recommendation system is a powerful tool to solve the problem of information overload,but traditional recommendation methods often face problems such as data sparseness and cold start,which leads to a decrease in the accuracy of recommendation.Comments contain a lot of rich and valuable information.How to more comprehensively dig out the characteristics of items and users' interest preferences from text comments,alleviate data sparseness and cold start problems,and achieve more accurate recommendations,has become a research Hot spot.In response to these issues,the main research contents of this thesis are as follows:Firstly,in view of the data sparsity problem in traditional recommendation methods,this thesis proposes a novel neural network recommendation model that can recommend items through user reviews.The respective comments of users and items are used to represent each user and item,and the recommendation problem is regarded as a text matching problem.For this reason,the CNN architecture and regression network are used to predict the matching score of the user-item pair to represent the user's rating of the item.The experimental results show that the model proposed in this chapter is better than the comment-based Deep Co NN and other neural network recommendation models,and can alleviate the problem of data sparsity to a certain extent.Secondly,for the item cold start problem,this thesis proposes a joint deep recommendation model,which is composed of two parallel neural networks to learn the low-level feature interactions of users and items respectively.Each network is further composed of two subnets.One of the sub-networks is used to use user and item comments,and the other sub-network is used to use metadata information and ratings to learn user preferences and item attributes.The potential features learned in each network are spliced together to generate potential feature vectors of users and items.The two networks are combined by introducing a shared layer at the top to learn the higher-level potential features obtained from the two networks and generate the final score.The experimental results show that compared with some existing recommendation models,the prediction performance of this model is better,and it alleviates the problem of cold start of items to a certain extent.Thirdly,the traditional recommendation method is to independently model the potential characteristics of user items based on static vectors,and does not consider the dynamic characteristics of user item interaction,which may affect the accuracy of the recommendation process.In response to this problem,this thesis proposes a recommendation model based on a multi-level attention mechanism.This model uses the neural network attention mechanism to jointly consider the fine-grained semantic information of user-item pairs,and uses specific interactive features based on comments to learn heterogeneous user item representations.Experiments with real data sets show that the model is better than the baseline method in terms of score prediction and ranking performance.