Research On Personalized News Recommendation Method Based On Deep Learning

Personalized news recommendation system can help users quickly obtain interesting content from massive news.The two main tasks in news recommendation are to obtain accurate user representation and news representation.Users' interests change with time,including long-term interests and short-term interests.The information contained in news is also divided into multiple categories,such as title,category,summary,etc,while the existing methods often learn the expression of news based on single category information.At this stage,the research focus of most recommendation systems is on how to improve the accuracy of recommendation.There is a lack of natural language interpretation of recommendation results,so it is difficult for users to know the reasons of recommendation results.If the recommendation result is given and the explanation of why the result is recommended is provided,it can not only improve the transparency of the system,but also improve the user's trust and acceptance of the system.Based on the above reasons,this paper mainly studies the following two aspects:(1)A method is proposed to fuse long and short-term user representation,and multi-featured news representation for personalized news recommendation.Firstly,a multi perspective learning method based on Collaborative attention mechanism is used to construct a news encoder to learn a unified news representation from the characteristics of news title,category and summary.Secondly,the improved news representation is used to further fine-grained learn the user representation in the user encoder based on long-term and short-term interest.The experimental results on real news data sets show that this method has significantly improved the accuracy compared with other recommendation algorithms.(2)An interpretable news recommendation method based on feature mapping and joint multi-task learning is proposed.Firstly,feature mapping is used to map unexplainable general features to interpretable aspect features,which eliminates the need for metadata in interpretable recommendation;At the same time,a joint learning model is used to balance the two tasks of accurate prediction and generation interpretation.While generating accurate recommendation results,the corresponding interpretation sentences are generated by using the implicit information between the two tasks,which realizes the accuracy and good explainability of recommendation.