Research Of Recommendation Algorithm Based On Opinion Mining And Deep Learning

With the rapid growth of information,recommender systems play a key role in alleviating information overload.It is well known that the recommender system is based on the users' historical behavior data and uses specific logic to filter and match the project of user preference.In recent years,the theoretical and technical applications of recommender systems have matured,recommender systems have been widely used in areas that require referral services,including e-commerce,online news and social media sites.At the beginning of the study of recommender systems,most of the research focused on more complex algorithms,more general framework,and did not pay much attention to experimental data.To this end,many recommender algorithms use user-item rating data as experimental data to verify the effect of the algorithm.The rating data do reflect the items that users' like,of cause,we can also predict the user's preferences based on this data.But rating data also exist some problems,the scope of rating data are fixed,and users cannot describe their evaluation of the items with more precise values,moreover,rating data cannot represent the true opinions of users completely because of some error operation.On the other hand,matrix factorization is one of the most popular method in the technology of collaborative filtering,user-item rating matrix will be decomposed into two matrices that contains latent factors,which is user matrix and item matrix respectively.Moreover,matrix factorization predicts rating by inner product of the two matrices.Matrix factorization technology is effective and widely applied,but the method has the limitation in predicting,matrix factorization will obtain a large error by using inner product to predict the complex user-item interactions in low-dimensional latent space.Aiming at the two problems mentioned above,this paper presents a recommendation framework that integrates various technologies.This framework aims to improve the experimental data and make predictions based on the values that are close to the user's real opinion,and the framework is also dedicated to break through the limitations of traditional matrix factorization methods.In recent years,it has been pointed out that the reviews provided by users were highly predictive and should be applied to the recommendation systems.At present,many review-based research have been proposed,they extracted topics,opinions and emotional polarity from reviews by utilizing text analysis and opinion mining.In order to make full use of the existed data and extract as much information as possible from dataset,this paper adopts a lexicon-based opinion mining technique to extract review elements and users' preferences from review data.At the same time,we utilize one deep neural network model to improve the deviation caused by matrix factorization in prediction process.This neural network model replaces inner product and learns the interaction function between users and items.Finally,we verify the validity of the proposed framework by using one public dataset in experiment.