Research On Movie Click-through Rate Prediction And Recommendation Based On Sentiment Analysis

With the rapid development of information technology and mobile Internet,traditional cinema films have come into public view in the new form of "Internet + Film".Film on players such as Tencent Video,i QIYI,Youku,Migu1 video,Watermelon video,Storm Player,Television Daquan and so on one after another,and film data increased and exploded.How to quickly find movies in line with subscriber’s interests and preferences from the huge amount of data,and how to maximize the merchants to attract new users and increase user loyalty has become a difficult problem.Personalized recommendation is the most convenient and effective way to solve these problems.It can integrate item attributes,user history behavior,and user characteristics(e.g.,user interest preferences)to tap valuable information from large amounts of data.On the one hand,it can help users efficiently access favorite movies,reduce repetition and loaverse information browsing;on the other hand,they can obtain more new users,increase user stickiness and improve the utilization of film data resources to achieve continuous growth of business goals.Traditional content-based and collaborative filtering recommendation algorithms have achieved excellent results in industry,but they are essentially recommendations based on user or interitem similarity calculations,highly dependent on historical information,and do not take full account of users’ real interest preferences.In addition,traditional recommendation algorithms rely on a large number of artificial features designed to automatically learn deep interaction features between users and items,and the effectiveness and scalability of the algorithms are limited.The paper adopts the recommendation by predicting the GBDT AFM model incorporating emotion analysis,that can avoid all problems above and satisfy the memory and generalization functions of the recommendation algorithm,the interactive information between continuous features and discrete features should be fully mined to improve the recommendation quality of movies.The implementation of the recommendation model mainly includes: 1.The sentence vector is calculated using the TF-IDF value as the weight of the keyword vector.The Stacking model of LR,SVM,MNB,Light TGBM and Fast Text is used to classify the sentiment of user comments and extract the sentiment value in comments.Comparing the Stacking model with each submodel,Staking shows that the classification accuracy of the Staking model is 91.65%,improved by 4.68%-13.31% over each submodel.Stacking model can fully extract emotional information from user comment data.2.The comprehensive score of user rating and user comment emotional value was calculated by viewpoint filtering method.Obviously,the overall score can better reflect the user’s real interest preference than the original score.For example,the user gave the original rating of 2 out of 2,with the comment "Adapted from a true story.I really admire the intelligence of the main characters!",the comprehensive score is "3.8".The user’s rating is obviously a bad review,but the comment information from the film is in favor of the attitude,the overall score of3.8 can better experience the real feelings of 47511486 users after watching the film.3.The GBDT algorithm was used to convert the statistical features of user rating,Douban rating,rating ratio,number of viewers,film length,proportion of comments and other features in the data set.Each decision tree is equivalent to a discrete feature,and the index of the leaf nodes of the sample decision tree is used as the value of the corresponding transformed feature,and finally a new feature data of one dimension is generated.The AUC value of the model on the test set reaches 96.05%,which indicates that the transformed feature can fully extract the original feature and denoise the original feature to a certain extent.4.The discrete features transformed by GBDT algorithm and the original discrete features are used as the input of AFM to realize the movie click-through rate prediction.When choosing a movie,users attach different importance to the content of the movie,the director and the rating of the movie website.By introducing attention mechanism layer,AFM model can learn the contribution degree of different interactive features to prediction results,which is more consistent with film data than FM model which treats different features equally.The final experimental results show that the combined model of GBDT+AFM is better than the combined model of GBDT+FM and the AFM model without GBDT feature.The loss of the GBDT+AFM model on the test set is significantly reduced compared to the loss of the GBDT+FM and AFM models,as low as 0.289730.In the AUC,accuracy rate,precision rate,recall rate and F1-score,the prediction results of the GBDT+AFM model have been significantly improved.The AUC value of the GBDT+AFM model reached92.91%,which increased 10.51% and 7.53% Compared with the GBDT+FM model and the AFM model.It can be seen from the loss value and the AUC value that the overall performance of click-to-record prediction with the GBDT+AFM model is better.From the perspective of accuracy,the accuracy rate of the GBDT+AFM model is 87.85% which is 7.24% higher than the GBDT+FM model and 6.33% for the AFM model.At the same time,the accuracy and recall rate of the GBDT+AFM model are also significantly improved,which shows that GBDT +The prediction results of the AFM model are more reliable.In general,the robustness and predictive performance of the GBDT+AFM model are better than the GBDT+FM model and the AFM model.Using the GBDT+AFM model for the prediction of the movie click rate has a better prediction effect.In addition,the recommendation result after integrating sentiment analysis is more in line with the user’s true interest preference than a movie recommended by a user who watched a movie as a positive sample.The former has a higher similarity between the recommended films and the highly rated films in the user history list.The GBDT+AFM model based on emotional analysis not only takes into account the user real interest preferences,but also is able to tap hidden information between continuous or discrete features,Which owned stronger generalization and learning capabilities.Film recommended with the model also has higher quality.