基于特征加权的神经网络集成及其应用

Feature Weighting based Ensemble of Adaptive Resonance Theory Networks and Its Application

泛化能力是机器学习关注的基本问题之一.特征加权是特征选择的更一般情况,它能更加细致地区分特征对结果影响的程度,往往能够获得比特征选择更好的或者至少相等的性能,已经成为普遍的提高学习器的泛化能力的方法之一.该文提出一种基于特征加权的神经网络集成方法FWEART,该方法通过自适应遗传算法的优胜劣汰机制为输入属性确定了特征权值,提高了集成中各个体Category ART网络的精度和差异度,从而提高了神经网络集成的泛化能力.在UCI标准数据集上验证了有效性后,FWEART被应用在地震序列类型预报上,取得了较好的预报效果.

Generalization ability is a principal issue in the field of machine learning. Feature weighting is a general case of feature selection, which has the potential of performing better （or at least similar） feature selection. This paper proposes a new ensemble method named FWEART （Feature Weighting based Ensemble of Adaptive Resonance Theory networks）, in which an adaptive genetic algorithm is used to conduct a search for the weight vector that can optimize the classification accuracy of the individual Category ART networks. Furthermore, the generalization ability of the ensemble is improved. Experiments on the UCI datasets show that FWEART has good generalization ability. Finally, FWEART is applied to predict the types of earthquake. The result is satisfactory.