Learning to Diffuse: Mechanism Design in Social Networks with Information Propagation Costs

Prior diffusion auctions assume that forwarding auction information is costless, making full propagation a dominant strategy. In networked markets and emerging multi-agent settings, however, diffusion may consume time, attention, bandwidth, or social capital, so agents must decide which neighbors are worth informing. We introduce the Learning Diffusion Mechanism (LDM), a repeated costly-diffusion extension of the Information Diffusion Mechanism (IDM). In each round, agents observe private valuations, choose whom to inform, and receive their IDM payoff minus a marginal cost per activated link. We show that truthful bidding remains weakly dominant under additive diffusion costs, which reduces the strategic problem to learning diffusion decisions. We then formulate the induced Bayesian diffusion game, establish existence of a stationary Bayes-Nash equilibrium, and derive a natural model-free policy-gradient update. LDM highlights how local propagation costs can shrink the realized auction network and provides a minimal framework for studying learned information diffusion in strategic networked markets.