Personalized News Recommendation Based On Gradient Boosting Decision Tree

With the continuous development of technology,people rely more and more on mobile phones and computers to get information.Users have different needs for miscellaneous news.Personalized recommendation to meet users' needs is a hot topic worth discussing.At present,convolutional neural network and self-attention mechanism are mainly used in the research of news recommendation field.This mechanism is mainly biased towards text features,and neglects more valuable attributes such as classification and time for the news field with strong timeliness and multiple topics,thus reducing recommendation accuracy.To solve this problem,this thesis proposes a personalized recommendation framework based on gradient elevation decision tree and logistic regression from the perspective of news features and users' interest preferences.The main research work includes:(1)A hybrid recommendation framework for news field is proposed.This framework uses Gradient Boost Decision Tree(GBDT)and Logistic Regression(LR)algorithm to represent candidate news according to its existing features,such as classification,title and content.According to the user's historical behavior,the user representation is combined and the input is combined into the click prediction representation to calculate the user's final preference for news,so as to obtain the personalized recommendation list.Compared with existing frameworks,the framework proposed by us improves multiple indicators such as recall rate and accuracy rate,and reduces the impact caused by data sparsity.(2)A news recommendation algorithm with balanced popularity and novelty is proposed.The algorithm classifies candidate news into hot database and new database according to users' click behavior.Hot database refers to the sequence with high click-through rate,while new database refers to the sequence with high click-through rate that has not been clicked.The popularity value of each news is given and fused with the recommendation list to obtain the personalized recommendation list for users.Compared with the existing recommendation algorithms,the proposed algorithm improves the recommendation accuracy.(3)A dynamic news recommendation algorithm integrating time factors is proposed.In order to meet the requirements of timeliness of news and improve the offset degree of users' interest in different time periods,the algorithm integrates two time factors into the news representation and user representation respectively by using the click time of users' access to news sequence and the stay time of users' access to news.Compared with the existing recommended algorithms,the proposed algorithm improves the prediction accuracy of user interest offset.This thesis uses MIcrosoft News Dataset(MIND),which is commonly used in the field of News recommendation algorithm,is used to demonstrate the superiority of the proposed framework and algorithm through experiments.In order to reduce the impact of data sparsity,improve the timeliness of recommendation,and focus on the changes of users' interests,this thesis expounds the specific recommendation process based on the principle of news recommendation,providing certain reference significance for future research in the field of news.