Research And Implementation Of Visual Interactive Recommendation Algorithm Based On Model Ensemble

With the development of Internet technology and cultural industries,more and more digital information,such as electronic goods,digital news,and online movies,is exploding in an exponential manner.It becomes very difficult for users to find their interested information from such massive data.The recommendation system can accurately predict the users' preferences based on the user's historical behavior information,help the user quickly find the information they are interested in,and greatly improve the efficiency of information dissemination.We propose the recommendation algorithm based on local model weighted ensemble through user clustering to solve the problem that a single model cannot capture different local features during the recommendation.Furthermore,we propose a local model weighted ensemble recommendation algorithm based on the selection of random anchor point pairs to solve the problem of overfitting for traditional models when the data are sparse.Based on the above-mentioned recommendation algorithms,the interactive visual movie recommendation system called RecVis is implemented to solve the black box problem of the traditional recommendation system.This system visualizes the user's portrait and recommendation process generated by the recommendation algorithm.We design several experiments with Douban movie datasets to demonstrate the effectiveness of the proposed algorithm.The case study of the visual movie recommendation system presented in this these shows that our system is useful and effective.The three main contributions of our paper are listed as follows:(1)Local model weighted ensemble based on user clustering: In order to solve the problem that a single recommendation model can not accurately capture different local features,we propose a local model weighted ensemble based on user clustering.This algorithm first adopts LDA topic modelling to train the feature vectors for each user through movie tag information to learn their tastes.Then the spectral clustering algorithm is used to get different user groups.The algorithm uses sparse linear model as the basic recommendation model and computes the local recommendation model for local groups and global recommendation model for all users respectively.The final ensemble recommendation model is obtained through linear weighted combination of the local and the global model.(2)Local model weighted ensemble based on the selection of random anchor point pairs: In order to solve overfitting of a single recommendation model for sparse training data,we propose a local model weighted ensemble recommendation algorithm based on the selection of random anchor point pairs.The algorithm uses LDA topic model and gradient boosting decision tree model to get user eigenvectors and movie eigenvectors.Multiple local training matrices are constructed through the selection of random anchor point pairs.The local recommendation models are trained from these local training matrices by sparse linear models.The final ensemble recommendation model is obtained through linear weighted combination of these local models.(3)Interactive visual movie recommendation system: In order to solve the problem of black box in the traditional recommendation system,we implement a web-based interactive visual movie recommendation system called RecVis.This system visualizes the users' portraits generated by the recommendation algorithm by showing the movie tag word cloud,the user theme radar chart,and the theme-level bubble chart,which improves the transparency of the entire system and enhances the user's understanding of the recommendation results.At the same time,the system allows the user to interact with the system by movie bubble cloud diagram.The user can customize the recommended model parameters,generate new recommendation results in real time.This system increases the user's control over the recommendation process,and makes users more involved in the recommendation.