Research On Federated Learning Method For CTR Prediction Scenarios

In the recommendation system,Click-Through-Rate(CTR)prediction is a key technology.Its task is to evaluate the probability of a user clicking on a certain recommended content.It usually recommends the most likely content for the user through a certain scoring rule,so as to increase the click-through rate of the recommended content,thereby increasing the revenue of the advertising system.However,the characteristics of a user on a platform are not enough to infer the user’s interest.Therefore,it is extremely challenging to infer the user’s interest when the user’s characteristics are not comprehensive.Fortunately,user behavior data usually exists on different platforms.For example,we will use Taobao Zhihu and so on,to fully utilize and mine user data on these platforms to obtain more comprehensive user information,morevover it can be more accurate to infer the user’s points of interest,finally increasing the click-through rate of recommended advertisements and increasing the platform’s revenue level.Unfortunately,user-related data has a high degree of privacy,due to user data privacy and security considerations and some data protection regulations such as the GDPR,user data on different platforms cannot be directly exchanged and collected.The existing CTR prediction models all require centralized collection of user data,so that the collected user data can be used to train the model.Therefore,if use user data on multiple platforms for CTR prediction tasks,it is not feasible to use the traditional CTR model.Fortunately,the proposal of federated learning can deal with the above problems very well.However,the existing federated learning methods in the field of CTR have problems that the scenarios are simple,that is,only scenarios where the models of all parties are the same are considered,and the problems that may exist in various actual business scenarios are not fully considered,such as different platform with different model structures to conduct training,etc.In addition,the existing methods do not effectively guarantee the privacy and security of user data.Aiming at the deficiencies of existing work in the CTR field,and considering the complexity and diversity of CTR scenarios,this paper proposes a new model Heterogenous Federated CTR.The model is based on the idea of knowledge distillation and can effectively deal with different scenarios in the CTR field.Even if the platform models are different,the model proposed in this paper can be directly used for joint training.Each platform only needs to exchange the output of their models,and exchange the knowledge of each platform with the help of a thirdparty central server through a unified output,so as to achieve the purpose of capturing the interest characteristics of users on each platform.The model training process only involves the exchange of model output information,and the user data of each platform does not leave the local at all,effectively ensuring the security of user data.Finally,in the experiment of this paper,a number of CTR scenarios are simulated,and the experimental results verify that the model proposed in this paper can effectively improve the performance of the original CTR model of each platform,that is,it is verified that the model can be used on the premise of ensuring user data security.It can effectively capture the interest characteristics of users distributed on different platforms.In addition,experiments in multiple scenarios in this paper also show that the model proposed in this article can deal with complex CTR scenarios and has potential application value.