• bayesian learning
• budget allocation
• online advertising
• reinforcement learning

The thesis aims to describe the structure of a framework useful for the allocation of budgets for online advertising campaigns. For each advertising campaign, will be taken into account the daily data related to the number of conversions and their cost. Subsequently, will be extracted a payout curve that describes the relationship between the conversion-cost variables. The curve will have a structure defined by a power law and model parameters will be extracted through Bayesian inference. The Bayesian approach, combined with the adoption of the Thompson-Sampling algorithm will allow us to address the problem of uncertainty and data sparsity with a mix of exploration and exploitation. Two methods for finding the optimal point will also be considered: a marginalistic approach and an approach based on the average value of a conversion. The two approaches will be studied and analyzed to find their strengths and weaknesses. Finally, is proposed a possible implementation of the framework designed with the support of PyMC3, a Python library specifically developed for probabilistic programming and Bayesian inference.