Parametric Bandits for Search Engine Marketing Optimisation

Expense optimisation for online marketing is a relevant and challenging task. In particular, the problem of splitting daily budget among campaigns, together with the problem of setting bids for the auctions that regulate ad appearance, have been recently cast as a multi-armed bandit problem. However, at the current state of the art several shortcomings limit practical applications. Indeed, campaigns are routinely divided by practitioners into sub-entities called ad groups, while current approaches take into account only the case of single ad groups: in this paper, we extend the state of the art to multiple ad groups. Moreover, we propose a contextual bandit model which achieves high data efficiency, especially important for campaigns with few clicks and/or small conversion rate. Our model exploits domain knowledge to greatly reduce the exploration space by using parametric Bayesian regression. Elicitation of prior distributions from domain experts is simplified by interpretability, while action selection is carried out by Thompson sampling and local optimisation methods. A simulation environment was built to compare the proposed approach to current state-of-the-art methods. Effectiveness of the proposed approach is confirmed by a rich set of numerical experiments, especially in the early days of marketing expense optimisation.