Learning to Sell Information

Abstract

Data buyers face a decision problem under uncertainty: they can augment their be- liefs by purchasing information from a seller. In the single buyer case, different types of buyers have varying willingness to purchase information, dependent on their prior information and beliefs. In the multiple buyer case, buyers produce negative exter- nalities on others when making the correct decision, and have different valuations on the importance of making the correct decision. We use deep learning to approach the problem of revenue-optimal mechanism design, formulating the setting as a learning problem and modeling it through a multi-layer neural network, which takes the re- ported prior beliefs or valuation of buyers, and outputs an allocation of statistical experiments that conveys information of varying quality to each buyer, together with their prices. We train the network using standard optimization pipelines to maximize the seller’s revenue and compared the performance of the learned mechanisms to the known analytical solutions.