| In the Internet era,personalized news recommendation is an important method to help people quickly find what they are interested in in a large number of news.The news recommendation process is usually divided into "recall" and "sorting" two stages.The recall stage is a rapid selection of a large number of candidate news,with the goal of containing as much news as possible that users are really interested in;In the sorting stage,the recalled news is sorted according to the degree that users may be interested in,and the news with the highest ranking is finally recommended to users.In most of the current news recommendation methods,it is often difficult to fully explore the potential semantic information of news in the recall stage,resulting in limited effects and easy to ignore users’ demands for news diversity.The ranking stage usually establishes a sequential recommendation task model with rich short-term dependence on the historical click news,which ignores the time diversity preference of users to news information and the dynamic migration of users’ interests.To solve the above problems,this paper studies news recommendation and proposes a personalized recommendation method based on deep learning.First of all,in the stage of recommendation recall,this paper proposes an "enhanced recall" model which combines hot recall,cluster recall and semantic recall based on pre-training.The hot recall model takes the time factor into account and forms a hot ranking according to the number of users’ clicks for recall.The clustering recall model makes clustering according to users,selects several users with greater similarity to target users in the class of target users and other classes respectively as the nearest neighbors,and recommends the news favored by these users to target users,which can make the news recommendation results more diverse.Based on the pre-trained semantic recall model,the pre-trained model Ro BERTa is used for news and user modeling,and the user behavior sequence and news information are fully integrated,which can effectively improve the recall rate.Therefore,the "enhanced recall" model,which integrates the three recall models,can increase the diversity of recommendation list news on the basis of improving the accuracy of recommendation.Secondly,in the stage of recommendation ranking,this paper proposes the Attention Mechanism and Temporal Diversity Recommendation Based on Conversation.The self-attention mechanism is used within the session and the attention mechanism is used between sessions respectively to learn the main reading interests of users within the session and the degree of correlation between the current session and the historical session,so as to capture the dynamic migration of users’ interests.The global and recent user interest relevance scores are constructed and calculated,and the final click score is obtained through the linear combination of the two,which can encourage the recommendation of news different from the recent click news and better meet users’ demand for the diversity of news information time.Finally,the personalized news recommendation algorithm model proposed in this paper was applied to Microsoft’s MIND public news dataset.Using a series of experiments,we selected the best combination of parameters for our model and compared it with the conventional recommendation model.The final experimental results showed that the recommendation effectiveness of the proposed method is superior to that of the compared methods,proving its feasibility and effectiveness. |
