融合子集特征级联预学习的封装方法研究

Feature selection algorithms in the machine learning domain can simplify the input of model, improve interpretability, and avoid dimensional catastrophe and over-fitting. In terms of selecting features on wrapper methods, the evaluation of models usually take the feature subsets filtered by the search algorithm as input directly, which leads to the fact that feature exploitation and evaluation of models is restricted by the feature reconstruction and fitting ability of the evaluation model. Moreover, the more appropriate feature subsets were limited to be discovered either. To solve the problems, a pre-learning wrapper method was proposed based on the cascade forest structure. Adding multi-level cascade forest between the search algorithm and the evaluation model, the model was arranged to transform the feature subset as high-level feature set, reducing the difficulty of recognition in the evaluation and improving the performance of feature subset. In contrast experiment, a variety of search algorithms and evaluation model combinations were evaluated on multiple datasets. The results indicate that the proposed method can reduce the number of selected features, while maintaining classification performance and the low coupling property of wrapper methods.