Ancillary Services in Targeted Advertising: from Prediction to Prescription

The role of optimization in some recent advances in data-driven decision-making

“…Our optimization-based approach to the problem offers more flexibility in terms of which response functions that can be fit (classical regression methods, tree-based methods, etc), which types of data that can be considered (discrete versus continuous), without needing to specify distributional assumptions on the data. These differences are further discussed in more details in [1] and [2].…”

“…The literature has considered clustering in offline learning before we proposed novel learning and optimization methods in [1] and [2]. As an early form of incorporating clusters in learning [10], the clusterwise regression problem considers outcomes of each observation to be described by a linear regression.…”

“…For example, while [18] propose a nonlinear formulation, [19] propose a compact mixed-integer linear formulation, and [20][21][22] propose a heuristic based on column generation for an integer linear formulation. In contrast to this literature, in [1] and [2], we move beyond linear regression and propose a more general clustered learning approach that can consider logistic regression, regression trees, classification trees, and others.…”

“…Our clustered learning approach, the Cluster-while-Estimate method (or CWE method), was developed in [1] for regression models and extended to classification models by [2]. The aim of this approach is to cluster observations (e.g., grouping similar products) whose outcomes (e.g., sales forecasts) can be predicted through a similar model.…”

“…In particular, we consider three parts (i) clustered learning methods, (ii) tree-based optimization methods, and (iii) model-free optimization methods. For each part, we will review some related literature, although we mainly focus on our own related research, particularly in [1][2][3], and [4].…”