NOVEL ACTIVE LEARNING BASED APPROACHES FOR BALANCING MULTI-OBJECTIVE MAXIMIZATION USING TRADE-OFF BETWEEN EXPLORATION AND EXPLOITATION

Abstract

Recently, active learning has gained significant interest in many applications where unlabeled data are abundant, however data labeling process could be expensive, hard to obtain, or time-consuming. Accordingly, active learning aims to choose informative and effective training examples for labeling. Most of the active learning work in literature focus on using active learning for minimizing generalization error for either classification or regression. However, active learning can be applied to decision making in cases where there is a utility function to be maximized while there exists some uncertainty in modeling such utility function. For example, we consider a business related problem of revenue maximization in case of unknown demand-price curve. In such problems, there is an inherent trade-off between exploration (to further improve the accuracy of the learning model), and exploitation (to utilize the model learned so far greedily to maximize the target utility). We seek to balance the exploration-exploitation trade-off in the revenue maximization with demand learning problem using two main approaches. In the first approach, we develop an active learning framework including several novel strategies that maximize the utility function while incorporating model uncertainty in different ways. For the second approach, we incorporate both objectives of revenue maximization and demand model learning into a novel hybrid utility function in several forms, and then we use price experimentation to maximize this hybrid utility function. We apply this approach, compared to the state-of-the art pricing methods, and this approach obtains promising results. Both of our proposed approaches seek to strike the balance between exploration and exploitation. In addition, both approaches achieve superior results in terms of the gained utility (revenue), achieving around 3-4% revenue gain over the competing methods in literature.