Data-driven decision model based on local two-stage weighted ensemble learning

To improve the decision performance using historical decision data, this paper proposes a data-driven decision model based on local two-stage weighted ensemble learning. The assessments of historical alternatives are collected from a multicriteria framework. For each new alternative, a set of its similar alternatives is determined from historical alternatives using the K-Nearest Neighbor technique, and then a set of base classifiers (BCs) is generated by the historical assessments. Based on ensemble error and diversity of BCs in predicting the similar historical alternatives of the new alternative, a local two-stage weighted ensemble method is developed to learn the optimal BC weights for the new alternative. Such a learning process not only considers the changes of BCs’ competence in facing different alternatives (instances) but also avoids falling into the dilemma of balancing the accuracy and diversity of BCs. By combining the continuous outputs of different BCs with the learned BC weights, the weighted ensemble outputs are obtained for the similar historical alternatives of the new alternative. Based on these outputs and the assessments of those similar historical alternatives on criteria, a linear optimization model is constructed to learn criterion weights. Using the learned criterion weights, the interpretable decision is performed. The advantages of the proposed decision model against four traditional decision models are validated by a real case study for the diagnosis of thyroid nodules. Thirty real datasets examine the competence of the proposed weighted ensemble method against mainstream ensemble methods and combination rules.