针对多元序列预测建模过程中特征选择问题,提出了一种基于数据驱动型高维k-近邻互信息的特征选择方法.该方法首先将数据驱动型k-近邻法扩展用于高维特征变量之间互信息的估计,然后采用前向累加策略给出全部特征最优排序,根据预设无关特征个数剔除无关特征,再利用后向交叉策略找出并剔除冗余特征,最终得到最优强相关特征子集.以Friedman数据、Housing数据和实际污水处理出水总磷预测数据为例,采用多层感知器神经网络预测模型进行仿真实验,验证了所提方法的有效性.%Feature selection plays an important role in the modeling and forecast of multivariate series. In this paper, we propose a feature selection method based on data-driven high-dimensional k-nearest neighbor mutual information. First, this method extends the k-nearest neighbor method to estimate the amount of mutual information among high-dimensional feature variables. Next, optimal sorting of all these features is achieved by adopting a forward accumulation strategy in which irrelevant features are eliminated according to a preset number. Then, redundant features are located and removed using a backward cross strategy. Lastly, this method obtains optimal subsets that feature a strong correlation. Using Friedman data, housing data, and actual effluent total-phosphorus forecast data from wastewater treatment plant as examples, we performed a simulation experiment by adopting a neural network forecast model with multilayer perception. The simulation results demonstrate the feasibility of the proposed method.
展开▼
