基于缓变特征学习的判别有序回归

Slow Feature Learning Discriminant for Ordinal Regression

有序回归是一种重要的机器学习范式,其目标是针对输出离散且有序的数据建立一个回归器以预测相应有序输出或离散类标。尽管现有的有序回归方法通过利用此类先验有序信息获得了比一般方法更优的性能。但是,并没有考虑缓变学习准则与有序回归的结合。本文通过缓变学习准则对每个样本类构建多个类内时间序列计算缓变类内散度矩阵,然后在有序约束条件的基础上根据线性判别准则寻找最佳投影进行有序映射,提出一种新的基于缓变特征学习的判别有序回归方法(Slow Feature Learning Discriminant for Ordinal Regression,SFLDOR)。通过在8个标准有序回归数据集上的对比实验表明,本算法在回归和分类性能上均优于使用普通类内散度矩阵的算法。

Ordinal regression is an important machine learning paradigm whose purpose is to predict the ordinal outputs or discrete labels by establishing an ordinal regressor .Many ordinal regression algorithms have been proposed with a better performance by u-sing this prior ordinal information .However , they do not consider the combination of slowness principle and ordinal regression . This paper first characterizes the slowness within-class scatter matrix with a number of within-class time series , and then establi-shes a slow feature learning discriminant for ordinal regression （ SFLDOR） by combining the matrix with an ordinal restriction . The experimental results with the eight standard ordinal regression data sets demonstrate that SFLDOR has a better regression and classification performance than the algorithm with general within-class scatter matrix .