Weighted Ensemble Incorporating New Strategies

The popularity of ensemble learning has propelled its application to diverse task. To improve its performance, weighted ensemble has become a widely adopted technique. However, existing methods consider the ensemble of weights directly on the variables, and seldom consider the feature selection of the variables before weighted ensemble. Unlike equally weighted ensembles, this paper proposes [Formula: see text] BCV-based learning of weighted ensembles. In order to utilize the data effectively, feature selection was conducted before ensemble learning. Contrary to the traditional methods, this study not only identifies features that significantly affect the response variable, but also improve the performance of ensemble learning. In addition, a detailed theoretical analysis of the estimated variance of prediction error based on blocked [Formula: see text] cross-validation is proposed. Finally, experimental analysis from simulated and real data sets prove that our proposed method is more effective and efficient. In particular, the proposed method was applied to tumor immunity, revealing that NKXB, BLB, the absolute value of CD4-positive cells, and the ratio of helper T-cells to cytotoxic T-cells as important features have a comparatively significant effect on gastric cancer, which provides a theoretical basis for clinical diagnosis.