Research On News Recommendation Algorithm Based On Deep Learning

In recent years,the Internet has been gradually integrated into every corner of life,so the recommendation system has been developed rapidly.The advantages and disadvantages of the recommendation algorithm are often measured by the click-through rate of online recommended products in the industry.Higher click-through rate can bring higher product revenue.Researchers and engineers have been working on algorithms to improve click-through rates,and CTR prediction models have been one of the important subjects of recommendations.From the perspective of CTR estimation,this thesis improves the accuracy of the ranking algorithm by optimizing the deep learning model,thereby improving the news recommendation effect.In addition,from the perspective of reinforcement learning,this thesis applies it to news recommendation scenarios and proposes a new recommendation algorithm.To solve the problem of using simulators to evaluate the reinforcement learning model exists big bias,this thesis proposes an off-lipe evaluation indicator.The main work and contributions of this thesis are as follows:(1)A sorting algorithm based on double cross deep network is proposed.This thesis proposes a Double Cross&Deep Network(DCDN)to improve the accuracy of news recommendation.This network is an improved algorithm based on the Deep&Cross Network(DCN)and can be applied to the field of CTR estimation,especially for news recommendation scenarios.The feature cross part of the model is improved to a double cross network,which enables it to focus on the cross modeling of certain types of features and other types of features,so that the model can learn the interactive information between features more efficiently and accurately.The deep network part of the model models the high-order nonlinear interactions of all features to improve the generalization of the model.Through a series of experiments on two real data sets,the proposed algorithm is proved to be superior to traditional machine learning algorithm and some deep learning algorithms.(2)A news recommendation algorithm based on deep reinforcement learning is proposed.This thesis introduces deep reinforcement learning that can solve sequence decision problems,applies the Deep Deterministic Policy Gradient(DDPG)algorithm to news recommendation scenarios,and uses generative adversarial learning to complement the advantages of the converged Q network and strategic network.This algorithm allows the recommendation system to take into account the dynamic changes in user interests and news content,and at the same time maximize the long-term benefits of continuous click behavior of users,rather than the short-term benefits of single click behavior.This thesis proposes an offline evaluation index QAUC for reinforcement learning models based on AUC.This indicator can simply and efficiently measure the advantages and disadvantages of reinforcement learning models,and facilitate comparison between models and evaluation of offline experiments.Experimental results on real data sets show that the algorithm used in this paper is superior to some non-deep reinforcement learning algorithms in news recommendation effect.