Data-driven decision model based on dynamical classifier selection

Abstract In the era of big data, large volumes of data have been accumulated in different fields. To help make decisions by using the accumulated data, this paper proposes a data-driven decision model based on dynamical classifier selection (DCS). Under the framework of multi-criteria decision making, historical data including individual assessments on criteria and overall assessments are collected and represented by interval-valued numbers. A set of base classifiers is selected and trained using historical data. For each new alternative, its similar historical alternatives are identified from historical data through using its prediction derived from each trained classifier. By using the predictive accuracy of each base classifier and the average similarity between the new alternative and its similar historical alternatives, a new DCS strategy is developed to select an appropriate classifier from base classifiers for the new alternative. The developed DCS strategy effectively avoids the subjective determination of the size of the local region in the field of DCS. Based on the similar historical alternatives determined by the selected base classifier, an optimization model is constructed to learn criterion weights from the individual assessments of the alternatives on criteria and the predicted overall assessments derived from the selected base classifier. Then, the explainable decisions are generated with the learned criterion weights. The proposed decision model is used to aid the diagnosis of thyroid nodules. Through its comparison with traditional decision model and three representative DCS methods, the advantage of the proposed decision model in balancing accuracy and interpretability is validated.