Prediction Confidence-Based Dynamic Selection And Weighted Integration

Ensemble learning is a successful learning paradigm and has gained considerable attention due to good generalization performance. Selection and integration of base learners are two key issues to construct an ensemble system. In this paper, a new kind framework of dynamic selection and weighted integration is proposed based on prediction confidence. In order to predict a sample, the base learners whose prediction confidences on the sample greater than or equal to a specified threshold are selected, and a final prediction is made based on the selected base learners with weighted integration. In weighted integration, the weight of a learner is computed based on its prediction confidence. Since the prediction confidences on different samples are different, the selected base learners as well as the weights are usually different for different samples. The computation of threshold is a key to the proposed framework. In classification ensemble, the threshold is computed via minimization of margin loss. In regression ensemble, the threshold is computed via minimization of prediction error. We analyze the rationality of the proposed framework in detail, and compare it with some other popular methods based on UCI classification tasks and short-term wind speed prediction, respectively. Experimental results validate its effectiveness.