Research On Critiquing Recommendation Algorithm Based On Variational Inference

With the continuous development of the Internet era,the amount of information on the Internet is increasing,and related problems like "information overload" are getting worse,as a result,it is difficult for users to get useful information quickly,so the recommendation system comes into being.The recommendation system can predict the needs of users and recommend the content they are most likely to like,which is used to solve people’s trouble of making choices from massive information.The core of the recommendation system is the recommendation algorithm,which can completely determine the performance of a recommendation system.Critiquing recommendation algorithm is a kind of dialogue recommendation algorithm,it can predict the project,but also can timely give the reason for the recommendation of the project,and provide users with a chance to recommit.The topic of this research is not only to improve the accuracy of recommendation items in the critiquing recommendation algorithm,but also to accurately judge the key word information of interpretation items.Specific research work is as follows:First,this thesis firstly formalized the definition and theoretical derivation of the model,and adopted the idea of combining variational inference and neural collaborative filtering to effectively enhance the characterization ability of the model.In addition,Bayes neural network was used to realize the model,and a robust model could be trained with less data.The loss function is designed from the perspective of maximum likelihood estimation,compared to the baseline model,the regression term in the loss function is eliminated,and the KL divergence is used to replace the autoencoder structure in the neural network,which not only makes the network structure more robust,but also greatly improves the convergence efficiency and recommendation accuracy.Second,it proposes to use the keyword combination information and attention mechanism to improve the recommendation effect.The attention mechanism learns the weight of keyword combination information through its influence on the prediction result,so that the model is more consistent with the real recommendation scenario.Thirdly,the pseudo-label training strategy is adopted in the model.Through semi-supervised learning,the unlabeled data is used to improve the performance of the model in the supervised process.The model trained on the labeled data is used to make predictions on the unlabeled data,the samples are screened according to the prediction results,and the high-quality data is re-input into the model for training.The model generates explanatory items through user and item coding,connects Bayesian neural networks to predict recommendations,at the same time,users will judge the explanation items they are not interested in,so that the recommendation system can re-predict the recommendation items.From the perspective of user interaction,users’ display preferences can be accurately mined,so that the recommendation system can predict the recommended items more accurately and improve the quality of recommendation.After several rounds of experimental comparison,the model has good performance in Precision,Recall,NDCG,MAP,R-Precision,and interpretation task index(F-MAP).In conclusion,the variational inference-based critiquing recommendation algorithm proposed in this thesis has a good performance in both recommendation and explanation terms.