摘要
传统深度森林模型由于局限性,在多粒度扫描特征转换阶段忽略了边缘信息,导致特征转换不充分;级联时将上一层类概率拼接到原始特征中,未考虑之前类概率向量的影响,最后投票过程忽视了子分类器权重。针对以上问题,提出一种特征重排序的深度森林(Reorder Feature Deep Forest,RFDF)算法,通过特征重排序,将较重要的特征排在中部转换出更有效的特征;级联阶段将之前层级生成的类概率向量之间的差作为增强特征与原特征拼接,进一步增强特征差异性,缓解网络退化现象。引入逻辑回归分类器,增加子分类器的差异性。最后结果由赋予权重后的子分类器投票得出。通过不同的数据集验证,表明该方法一定程度上有效,在高维数据集上表现更加明显。
In the traditional Deep forest(DF),its own limitations will make the feature conversion stage of multi-granularity scan ignore edge information,causing insufficient conversion.When cascading,the class probability of the previous layer is spliced into the original feature,without considering the influence of the previous class probability vector,and the final voting process ignores the weight of the subclassifier.In view of the above problems,a feature-reordered Weighted Deep Forest is proposed.Through feature reordering,arrange the more important features in the middle to transform into more effective features.In the cascade stage,the difference between the class probability vectors generated by the previous level is used as the enhancement feature and the original feature stitching,further enhanced feature difference,mitigating network degradation.And introduce a logistic regression classifier to increase the difference of sub-classifiers.The final result is obtained by voting after the weighted sub-classifier.Through the verification of different data sets,this method is effective to a certain extent,and it is more obvious on high-dimensional data sets.
作者
周博文
皋军
ZHOU Bo-wen;GAO Jun(College of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212032,China;College of Information Engineering,Yancheng Institute of Technology,Yancheng 224000,China)
出处
《软件导刊》
2021年第9期7-13,共7页
Software Guide
基金
国家自然科学基金项目(61772198)。
关键词
深度森林
特征重排序
增强特征
加权森林
deep forest
feature-reordered
enhanced feature
weighted forest