A Click-through Rate Prediction Algorithm Based On Sparse Behavior Data

With the development of the Internet,big data and cloud computing technology,peo-ple live in a sea of information.This rich information meets the needs of users,but the large amount of data also makes them suffer.On the one hand,located the useful information is inefficient.On the other hand,it is difficult to find the information really needed from the massive amount of information.Click-through rate prediction aims to achieve accurate matching between users and products,to recommend the content that they are most likely to be interested in,and to eliminate the impact of information overload.Therefore,click-through rate prediction has always been a hot research topic in the field of user behavior analysis.However,user behavior data is often sparse and exists in a long-tail distribution,mak-ing it difficult for the model to accurately understand user intent.In addition,due to the existence of data islands,the sparse nature of user behavior data is further aggravated.Although data sharing can improve the accuracy of user behavior understanding,violent and direct data sharing will leak user privacy information,resulting in damage to the inter-ests of multiple parties.To address the above two issues,this thesis studies their problem from the perspectives of user behavior and data isolation,and proposes two click-through rate prediction algorithms that can solve the problems of long-tail distribution and isolated data island.The main contributions of the thesis are as follows:· Aiming at the problem of sparse behavior,this thesis proposes a click-through rate prediction algorithm based on behavior layering.The existing click-through rate prediction algorithms design model with strong fea-ture extraction capabilities from the perspective of model architecture,but rarely consider how to improve user behavior prediction capabilities from the perspec-tive of data sparseness.To solve the problem of inaccurate predictions caused by sparse user behavior data,the thesis designs a user behavior layering strategy.The user population is divided into different levels of subgroups according to the data sparseness.The click-through rate prediction model selects different behavior tar-geting strategies for different levels of subgroups.The experiment results verify the effectiveness of the user behavior layering method.· Aiming at the problem of data security sharing,the thesis proposes a click-through rate prediction algorithm based on federated learning.The phenomenon of isolated data island is widespread,and the security of private information has at-tracted more and more attention from the society.Isolated data island exacerbates sparsity.Simple and casual data sharing will leak user privacy and cause serious consequences.In order to realize safe user behavior sharing,the paper proposes a click-through rate prediction algorithm based on federated learning,using privacy protection technology to allow each client to update the local model parameters,and only upload the encrypted gradient to the central server for safe aggregation.Ex-periment results demonstrate that the method can accurately realize user behavior orientation while ensuring the security of data sharing.