Research On Rating Prediction Methods And Key Technology In Recommender Systems

The rating prediction in recommender systems can predict user missing rating value through the known historical rating records.It can be widely used in e-commerce,tourism,social networking and other fields,with considerable application prospect and research value.However,the existing rating prediction models are faced with such problems as lack of labeling rating data and long-tailed distribution,which lead to low prediction accuracy,and cannot meet the needs of the development of recommender systems.Therefore,it is urgent to improve the performance of rating predictions and ensure the accuracy of predictions.Based on deep learning and machine learning technologies,this paper made a rich research on rating predictions by using representation learning to improve collaborative filtering,learning global item features to make robust predictions,integrating social trust and sparse autoencoder model for predictions,and using the transfer of rating knowledge to realize cross-domain predictions.Firstly,in view of data sparsity problem and difficulty in extracting interactive rating information,a collaborative filtering rating prediction method based on entity co-occurrence representation learning and a collaborative filtering rating prediction method combining global and local rating representation learning are proposed.Based on the principle of word co-occurrence,the semantic information of interactive ratings is extracted from two perspectives of entity co-occurrence and entity rating,respectively.By decomposing entity and rating co-occurrence matrices,the embedding features are represented,and the least square method is used to build the training model for characterizing and predicting user rating values.Secondly,in view of the local defects and weak robustness of existing rating prediction methods,a global item rating prediction based on robustness constraints is proposed.The outliers in the data sample have an important influence on the performance of rating predictions.This method first proposes a robust constraint strategy to eliminate the outliers in the rating sample to avoid the bias influence of the outliers on the prediction results.According to the different features among items,the difference weight and compensation coefficient of items are proposed to improve the rating benchmark predictor,and then user preferences are learned and the final predictions are made.Thirdly,in view of the problems of single recommendations and low prediction accuracy of the traditional rating predictions,a rating prediction method based on social trust ensembling and deep autoencoder learning is proposed.By mining and extracting the rating information of users' social trusting neighbors,the multi-source rating data is obtained to characterize user preferences,which are incorporated into the matrix decomposition.Based on the obtained multi-source rating data,the regularization and sparse constraints are added to the deep sparse autoencoder model to realize the rating prediction results.Then,a cross-domain rating prediction method based on the latent feature transfer of partially overlapping entities is proposed to solve the data sparsity problem and domain heterogeneity.In the proposed method,the latent features of entities are extracted by cross-domain matrix decomposition,and then the user latent features in the source and target domains are aligned by generative adversarial network model.Based on aligned user features,the cross-domain item features are also aligned.In addition,an N-step cross-domain random walk algorithm is designed to build cross-domain user relationship diagram and calculate user similarity across domains.According to the obtained user similarity,a cross-domain rating prediction model based on similarity constraint is proposed.Finally,in order to evaluate the performance of rating prediction methods proposed in our paper,the comparison experiments on Epinions and Movie Lens 25 M datasets were carried out,and analyses were made through mean absolute error,root mean square error,accuracy,recall and convergence indicators.