Distributed representations of bids in lowest unique bid auctions

In this paper, we study multi-item lowest unique bid auctions (LUBA) with resubmission under budget constraints. A representation of bids that can perceive the rich strategic meaning in the world of multi-item LUBA with resubmission is learned by an efficient method-Skip-gram model. We created and maintained a dataset which contains the relevant information of LUBA, including the winning bid combination, budget constraints, and the number of participated bidders. An Android-based application is developed and contributes the construction of the dataset. The quantitative analysis displays that the representation cluster can reflect similarities in terms of numerical information, cultural and aesthetic preference, and other bidder's behaviors. The learned representation can serve as a guide for bidders who seek to maximize their payoff and feed into a sequence generation model, such as recurrent neural network, to produce the bids combination.