Research And Implementation Of Revenue Optimization Algorithm In Real-Time Bidding System

In recent years,with the rapid development of Internet software and hardware and the popularization of various devices and applications,the Internet has been integrated into every aspect of daily life.Meanwhile,it also has a far-reaching impact on all walks of life.The real-time bidding trading mode in the Internet advertising market has emerged and been constantly developed and improved.Based on the large amount of data and personal information generated and accumulated by users' online behaviors,real-time bidding realizes accurate and efficient advertisement delivery by means of multi-field technologies such as machine learning,information retrieval and optimization theory,and has become an important means to make profits by Internet traffic.Due to its promising development prospects and rich application technologies,real-time bidding has attracted more and more researchers' attention.In this thesis,several key problems are studied from the perspective of demander,including click-through rate prediction,budget cost management,bid strategy optimization,and the corresponding algorithm is proposed to optimize the overall revenue of demander.The research results of this thesis are as follows:(1)To improve the accuracy of click-through rate prediction,The actor-critic reinforcement learning framework was first applied to the click-through prediction task.In this thesis,the target and concept of click rate prediction task are reasonably modeled,and a click-through rate prediction algorithm based on actor-critic framework is proposed,and neural network is used to learn the policy and value function,the framework's training and testing procedures are defined.(2)In order to improve the efficiency and smoothness of budget spending,this thesis proposes a budget allocation and management mechanism,which includes two parts: the budget allocation strategy based on portfolio theory and the impression request filtering mechanism.Budget allocation strategy considers revenue and risks these two factors at the same time and can ensure certain expected benefits while minimizing the risks of the allocation plan,so as to obtain more stable benefits.The impression request filtering mechanism filters out low quality ad requests by setting a threshold for the predicted click-through rate of requests,thus improving the use efficiency of the budget.Also,budget dynamic adjustment between slots is introduced to make the budget use more sufficient.(3)Aiming at the optimization of bidding strategy,this thesis modelled this problem as an optimization problem with budget constraints,by solving this problem this thesis proposes a slot-wise nonlinear bidding strategy,which combines the budget allocation and management mechanism,and can adjust the bidding strategy according to the actual budget spending in the advertising campaign.This bidding strategy only bids for high-quality impression requests and skip the low-quality impression requests,so as to obtain more reliable bidding results.