Research Of Context-based Deep Hybrid Recommender System

The sparseness of user-to-project scoring data is one of the main factors in the deterioration of the recommended system.In order to deal with sparseness problems,some recommendation techniques consider auxiliary information to improve the accuracy of scoring prediction.When scoring data is particularly sparse,document-based modeling approaches improve accuracy by using textual data such as reviews,abstracts,or synopses.However,due to the inherent limitations of the term bag model,they still have difficulties in effectively using the contextual information of the document and can only understand the document shallowly.ConvMF algorithm model can capture the context information of the document and further improve the prediction accuracy of the score.However,the convolutional neural network in ConvMF ignores the feature's location information when analyzing the text,and loses the information strength of the same feature.In order to solve the problems in the above models,this paper deeply researches the convolutional network and proposes a novel context-based deep hybrid recommendation algorithm(ConvMF with Segmeng-Max Pooling,ConvMF-S).The content of this article is mainly divided into the following sections:First of all,this paper introduces the current research status of recommendation systems,and introduces various algorithms in the recommendation system,especially the most classic collaborative filtering technology in personalized recommendation systems.After fully understanding the principles of various algorithms,the advantages and disadvantages of these algorithms were analyzed in depth.Then introduced the application of deep learning in the recommendation system.Secondly,this paper deeply studies the application of convolutional neural network in natural language processing,improves the existing ConvMF model,and proposes a new model ConvMF-S.Again,this article has conducted an in-depth study of the optimization of the ConvMF-S model.First,the CNN embedded layer is initialized by pre-training the word vector model,and then the model is further optimized by the coordinate reduction algorithm and RMSProp.Finally,this paper validates the improved model algorithm through relevant experiments,and compares it with other algorithms from various aspects,and analyzes and evaluates the experimental results.