Application Research Of Machine Learning In Combinatorial Auction Mechanism

Combinatorial auction sells bundled items by public bidding,allocates items to users according to allocation rules,and calculates users’ payment according to payment rules.Then providers make profits during the transaction process.Combinatorial auction can effectively improve the efficiency of item allocation and is widely used in modern auctions.However,combinatorial auctions mainly consider how to allocate items to users,and seldom consider the provider’s revenue.The provider’s revenue is mainly affected by the following three aspects.First,the valuation of the same item bundled by different users is unevenly distributed,which leads to the user with a higher bid to win the auction.In fact,the winning user’s payment is often lower than his bid,so that the provider’s revenue still has room for improvement.Secondly,users are selfish,and they will strategically lie about the price to manipulate the auction to improve their own utility,which leads to the user’s low payment and damages the provider’s revenue.Finally,the number of items for sale is closely related to the provider’s revenue.Multi-item multi-user combinatorial auction can significantly improve the provider’s revenue.However,most of the existing research focuses on single-item auction.Considering the above problems,this paper aims to study the use of machine learning algorithms to improve the revenue of providers in combinatorial auctions.The main contents can be divided into the following three points.(1)Research on combinatorial auction mechanism based on random walk algorithm.Aiming at the problem of low revenue of providers in combinatorial auction,considering the uneven distribution of user valuation,a combinatorial auction mechanism based on random walk is designed to maximize the revenue of providers.Firstly,combined with the concept of virtual valuation,the random walk algorithm is used to search the optimal parameter in the parameter space,and the user valuation is linearly transformed by the parameter.Then,the virtual valuation after linear transformation is used as the input of the Vickrey-Clarke-Groves(VCG)mechanism to determine the users who win the auction and calculate the corresponding payment.Theoretical analysis shows that the proposed mechanism satisfies the incentive compatibility and individual rationality.The experimental results show that the proposed mechanism can effectively improve providers’ revenue.(2)Research on combinatorial auction mechanism based on genetic algorithm.Considering that the number of parameters of virtual valuation increases exponentially when the number of users and items increases,the virtual valuation model with reduced parameters is studied,and a combinatorial auction mechanism based on genetic algorithm is designed.The user valuation is sampled from the prior distribution and mapped,and the virtual valuation is used as the input of the VCG.The winner determination problem is modeled as a constrained integer programming problem,and the genetic algorithm is used to search for the global optimal parameters.Simulation experiments show that the proposed mechanism can effectively improve the revenue of providers while maintaining incentive compatibility and individual rationality.(3)Research on combined auction mechanism based on deep learning.Since the user’s valuation space is complex and difficult to estimate,we explore the use of deep learning tools to automatically design combinatorial auctions to maximize provider revenue.Formulate optimal auction design as a constrained learning problem,modeling network and design mechanism based on residual attention mechanism.Input the user’s valuation and preference information for any item into the learning framework,design a network to automatically solve the optimal mechanism,and output item distribution information and user payment results.Experimental results show that the designed mechanism can achieve better results in combined auctions with different items and numbers of users.This paper studies the revenue of providers in combinatorial auction.From the perspective of mechanism design and machine learning,the formal modeling of combinatorial auction is carried out,and the network structure of item allocation and user payment is designed.The results show that the mechanism designed in this paper can effectively improve the revenue of providers.