Table of Contents
SomeIn this approach, we first build a set of models \(M_{11}, M_{12}, \ldots\) for each action allowed in the initial state, using profiles of customers who received this action in the past and with the target label defined over a long period of time. This label is a direct estimate of the expected return. We then build a second set of models \(M_{21}, M_{22}, \ldots\) using customer profiles with features that indicate which action was taken first, and so on. In other words, we optimize actions sequentially using the fact that “good” customer trajectories start with “good” initial actions for given customer characteristics.thing
