Regression Tree for Bandits Models in A/B Testing

Emmanuelle Claeys, Pierre Gançarski, Myriam Maumy-Bertrand, Hubert Wassner

Abstract: In the context of Web A/B testing, dynamic assignment of traffic aims to promote the best variation (A or B) as quickly as possible. However, dynamic assignment is difficult to use when the difference between A and B affects the visitor differently according to his/her personal characteristics and his/her history (number of visits, navigation on the website ...). In this paper, we propose a dynamic assignment strategy based on a visitor segmentation determined automatically from the visitors navigation and characteristics.