Personalized Recommender System Based On Heterogeneous Multi-Level Feature Information

With the development of society and Internet technology,the number of users and the number of items on the website have explored sharply.This makes it more and more difficult for users to observe the items,which are existed in the numerous objects.Therefore,the personalized recommender systems emerge at the historic moment.In the recommender system,the rating of the user on an item is a typical explicit feedback.A specific rating system can be utilized to select different rating values to illustrate the user’s preference for different items.However,not all users give objective ratings to items,which will result in a sparse user-item rating matrix,therefore,it is relatively hard to recommend items to users.Fortunately,the additional auxiliary information can alleviate the problem of data sparsity in recommender system,thereby improving the accuracy of rating prediction.This paper mainly utilizes a comprehensive and in-depth mining of a large number of attributes of items as context information to alleviate the sparseness of the rating matrix.The proposed models utilize a variety of deep learning neural network algorithms to extract heterogeneous multi-level features,and then seamlessly connect the extracted features to the recommender system.This paper mainly includes the following three parts:1.Aiming at the problem that the auxiliary information of the traditional recommender system is single and the neural network structure is relatively limited,this chapter proposes a multi-category context-aware recommender system(Visual convolutional matrix factorization,VConv MF),which integrates visual feature and textual feature to improve the accuracy of rating prediction in recommender system.Specifically,the pre-trained word embedding model is firstly used to convert the user’s comment text into a dense digital matrix,that is,the natural language is converted into machine language available to the machine.Then,the deep convolutional neural network is used to extract the features of the objective item images,which can easily extract the corresponding features from the shallow layer to the deep layer for further fusion.Then,the textual information and visual information extracted from the above two parts are cascaded and embedded into the Probabilistic matrix factorization(PMF)model to predict the unknown ratings of items in the matrix.Through the experimental simulation of three real-world datasets,and compared with other advanced recommendation algorithms,the simulation experiment results show that the proposed VConv MF model has good results.2.As for the problems in convolutional neural network mentioned in the process of text processing,this chapter proposes a recommender system algorithm(Recurrent convolutional matrix factorization,RConv MF)based on the recurrent convolutional neural network model.To further improve the performance of the context-aware recommender system,the recurrent convolutional neural network is utilized to process textual information and the convolutional neural network is utilized to process image information.Then,the textual information and visual information extracted from the above two parts are cascaded and embedded into the probabilistic matrix factorization(PMF)model to predict the unknown ratings in the rating matrix.Through the experimental simulation of three real-world datasets,and compared with other advanced recommendation algorithms,the simulation results show that the proposed VRConv MF model has good results.3.In view of the fact that the auxiliary information of the classic recommender system is static and incomplete,this chapter proposes a recommender system algorithm(Trailer-inception probabilistic matrix factorization,Ti-PMF),which utilizes the dynamic visual information extracted in trailers as item latent models.First of all,in order to prevent the high coupling of the images,the model needs to evenly extract still frames from the video clips.Then,all the still frames in the same video clip are transformed into corresponding description text through the neural image caption generator model.The obtained description texts are converted into textual features through the recurrent convolutional neural network model,and then embedded into the PMF model.Due to the numerous models,the visual and textual model are pre-trained by using multiple datasets,followed by transfer learning.Finally,the unknown ratings of users in the rating matrix is predicted.Through the experimental simulation of three real-world datasets and the comparison with other advanced recommender schemes,the simulation results show that the proposed Ti-PMF model has good results.