A Study On Dynamic Recommendation Algorithms In Recommendation Systems

With the rapid development of Internet technology,human beings have entered the era of big data,the problem of "information overload" has become one of the urgent problems to be solved.Recommendation system,which is one of the technologies to solve the problem of“information overload”,has been widely used in Internet applications.Traditional recommendation technologies which based on statistical learning and deep learning cope with the dynamic change of item popularity and the update of recommendation candidate set by periodically updating the model,which can not update the recommendation priority in time according to the change of item popularity and the cold start process for new added items can not be completed quickly.Dynamic recommendation algorithms(such as Bandit)can solve the above problems to some extent.However,the performance also needs to be improved,that is because:1)Their ability of models is limited,for example,LinUCB utilizes a linear model to fit the user's interest on a particular item,the representational ability is limited,which caused the limitation of the model.2)They do not take into account the heterogeneity of distribution for users'features,the recommendation performance is not good.In view of the above two problems,this paper chosen news recommendation as a specific scenario for dynamic recommendation algorithm study.Based on a large-scale and real-world online news system user behavior log,we measured and modeled the dynamic change of news popularity in the news recommendation system,the pattern of news for added or removed from candidate set,and observed the distribution of user features.Based on the observation results,we proposed two algorithms to solve the above two problems and evaluated our proposed algorithms based on the real-world data.The main contributions are as follows:(1)Aiming at the problem of poor representational ability of existing models,this paper proposed to use neural network to model the relationship between the user feature and the expected reward,besides,we solved two difficult problems of neural network updating and loss function selection.Specifically,first of all,in order to solve the problem that the online training of neural network is difficult to converge in the case when the samples are unbalanced,we proposed a user feedback aware training method:feeding the neural network with different times according to different user feedbacks,compared to the traditional training method,this method achieves nearly 40%performance gain.Secondly,this paper modeled the recommendation problem as regression,classification and policy gradient.We tried various loss functions,including classification,regression and policy gradient.Experiments show that under a reasonable configuration,using the loss function of policy gradient,our algorithn achieves 2.1%performance gain compared to LinUCB algorithm,which proves the performance of the algorithm.(2)In view of the fact that the traditional Contextual-Bandit algorithm does not consider the heterogeneity of users' features,this paper proposed an innovative hierarchical recommendation algorithm which is aware of user features.This algorithm can dynamically identify the categories of users,and then dynamically match the appropriate recommenders according to the categories of users to obtain the best recommendation performance.Experiments show that the proposed algorithm achieves 3.3%performance gain compared to the traditional LinUCB recommendation algorithm,which proves the performance of the proposed algorithm.The study on the Contextual-bandit recommendation algorithms in this paper,improved the performance of the current mainstream dynamic recomnendation algorithms,has certain theoretical values and application values.