Cross-domain decision making with parameter transfer based on value function

Abstract

Decision models based on value function can help learn the preferences of decision makers using historical data. However, inconsistency of the preferences learned from data with the real preferences of decision makers may occur. Suppose decision data or scenarios in the target domain are insufficient to guarantee the consistency of learning preferences, and there exists one similar source domain with sufficient effective data. Then, knowledge from the source domain can help learn preferences in the target domain. Following this idea, this paper proposes a cross-domain decision making method using a parameter transfer strategy with homogeneous and heterogeneous criteria. A decision model based on the value function is constructed using a generalized additive model to ensure the interpretability of the model. The marginal value functions are adjusted using Sturm’s theorem to keep the monotonicity of the learned marginal value functions on the criteria. Then, the parameter transfer strategy is adopted to transfer the obtained value functions to help learn the real preferences of the decision maker for the decision problem where preferences in historical data may not be consistent. The effectiveness of the proposed method is validated by solving the problem of diagnosing breast lesions in a hospital.