实例加权类依赖Relief

Relief算法是一个过滤式特征选择算法，通过一种贪心的方式最大化最近邻居分类器中的实例边距，结合局部权重方法有作者提出了为每个类别分别训练一个特征权重的类依赖Relief算法（Class Dependent RELIEF algorithm：CDRELIEF）.该方法更能反映特征相关性，但是其训练出的特征权重仅仅对于衡量特征对于某一个类的相关性很有效，在实际分类中分类精度不够高.为了将CDRELIEF算法应用于分类过程，本文改变权重更新过程，并给训练集中的每个实例赋予一个实例权重值，通过将实例权重值结合到权重更新公式中从而排除远离分类边界的数据点和离群点对权重更新的影响，进而提高分类准确率.本文提出的实例加权类依赖RELIEF （IWCDRELIEF）在多个UCI二类数据集上，与CDRELIEF进行测试比较.实验结果表明本文提出的算法相比CDRELIEF算法有明显的提高.

Instance Weighted Class Dependent Relief

The Relief algorithm is a filtering feature selection algorithm that maximizes the instance margins in the nearest neighbor classifier in a greedy manner. Combined with the local weight method, the authors proposed a Class Dependent RELIEF (CDRELIEF) algorithm that trains one feature weight for each category. This method can better reflect the correlation of features. However, feature weight vector are only effective for measuring the correlation of features to a certain class, and classifying them in actual classification. In the actual classification, the classification accuracy is not high enough. In order to apply the CDRELIEF algorithm to the classification process, this study changes the weight update process, and assigns an instance weight to each instance in the training set. By combining the instance weight value into the weight updating formula, the influence of data points far from the classification boundary and outliers on weight updating is excluded, thereby improving the classification accuracy. The Instance Weighted CDRELIEF (IWCDRELIEF) algorithm proposed in this study is compared with CDRELIEF algorithm on multiple UCI 2-class datasets. Experimental results show that the algorithm proposed in this study has significantly improved the CDRELIEF algorithm.