Research Of News Recommendation Algorithm Based On Deep Learning

With the development of public security information construction,a large number of public security data have been collected in police work.As a part of the public security data,news data has become an important window which can help the state and police understand social dynamics and social information.How to accurately obtain the required information from these massive news data through the recommendation algorithm,which can provide support for the guidance and monitoring of public opinion,has become a popular research direction.Currently,content-based recommendation algorithms have become one of the most widely used recommendation algorithms due to their excellent performance,and the accuracy of recommendation is further improved by combining with deep learning technology.However,the recommendation results of existing recommendation algorithms are susceptible to data sparsity,and many of the methods of integrating deep learning only focus on the learning of text content,and pay fewer attentions on to the analysis of user behavior.In addition,the new news information added in the actual scene will cause the item cold-start question of the recommendation system,which also has a certain impact on the recommendation accuracy.Therefore,the following researches are carried out in this thesis.Aiming at the data sparsity problem of existing recommendation algorithms,News Recommendation Model based on Behavior Embedding(NRMBE)is proposed by introducing Skip-gram network.Firstly,the user's browsing data is processed aiming to obtain serialized data.Then,in order to improve the learning ability of Skip-gram network,a global variable is added to optimize the encoding method of the central sequence of the network.And the improved Skip-gram model is used to learn the user's browsing behavior habits.Finally,the recommendation for target user can be calculated by using the cosine similarity formula.The experimental results show that after the sparse browsing data is processed by the improved Skip-gram network,the recommendation accuracy of the recommendation model has been improved to a certain extent.In order to solve the item cold-start problem,Cold-start Recommendation Model based on Paragraph Embedding(CRMPE)is proposed.Firstly,the Doc2 vec network is used to calculate the N most similar texts between the original news text and the newly added news text.Then,these texts use the improved Skip-gram network to calculate theencoding based on the user's browsing behavior,and on this basis,the new text is numerically calculated to obtain the encoding based on the user's browsing behavior.Finally,new items can be recommended by using the cosine similarity formula.The experimental results show that the accuracy and coverage of the cold start recommendation model are improved by the fusion of the Doc2 vec network and the improved Skip-gram network.Finally,based on the above fusion model,a prototype system of news recommendation is designed and implemented.