Adaptive Linear Contextual Bandits For Online Recommendation

In recent years,the rapid development of big data technology has made digital information resources grow exponentially.Various types of digital information are numerous and complex,and are flooding the entire Internet space.How the user selects the information needed from the massive data and how the merchant can accurately sell the product to the users who need it,these requirements make the recommendation system more and more important.The personalized recommen-dation system based on deep learning of big data is also Flourish.At present,a large number of Internet activities are completed through real-time online.More and more users choose to listen to music,shopping,and watching movies online.The real-time online recommendation system attracts much attention in the field of recommendation.It requires the recommendation system to respond quickly to user feedback,create user portraits,explore and explore user interests,and recommend favorite items for users.The context-based multi-arm gaming machine algorithm originated in gam-bling,and its core is to solve the problem of how to make choices.In recent years,more and more algorithms have been applied to the online recommendation field and have achieved good results.The algorithm can construct user portraits,and continuously explore and maintain user portraits.Through continuous learning in the recommendation process,the algorithm can learn more and more about the user.However,in the process of making a recommendation for the user,the algorithm uses the interest preference features stored in all user portraits for any item,and ignores the selective use of related interest preferences for special items.The introduction of too much noise in the recommendation process reduces the accuracy of the recommendation.Therefore,in the research work of this thesis,we proposed an adaptive con-textual bandits algorithm(AdaLinUCB).In the recommendation process,the algorithm first uses a user interest filter matrix to filter the user interest prefer-ence parameters adaptively for each specific item to obtain a new user interest preference.Then,for each candidate recommendation item,score prediction is performed using the corresponding filtered user interest preference parameter.Subsequently,the item with the highest score prediction in all items is recom-mended to the user,and the user will make a real rating feedback on the item.The final recommendation system uses the obtained rating feedback to update the user interest preference parameters and the user interest filter matrix so that a more suitable item can be recommended for the user in the next recommenda-tion.In this recommendation process,we propose an alternative gradient descent method to update and learn parameters,and use the online sliding window model to manage the user recommendation history,and strive to influence the recom-mendation update efficiency during the online update process.Minimizes and ensures that the recommendation accuracy rate is improved without sacrificing real-time online recommendation efficiency.We conducted a large number of experiments on two public datasets.Com-pared with a large number of benchmarking algorithms,our proposed adap-tive contextual bandits algorithm achieved good experimental results and even achieved feedback on user cumulative scoring.Improved the effect of 15%.At the same time,we also conducted comparative experiments on real-time recommen-dation efficiency.Our algorithm has a good time efficiency performance while improving the recommendation accuracy rate.After visualizing the user interest filtering matrix,the deep learning process and results of the algorithm were found to be interpretable,thus verifying that the adaptation is a contextual multi-arm gamer algorithm.