基于稀疏学习的鲁棒自表达属性选择算法

FEATURE SELECTION ALGORITHM WITH ROBUST SELF-REPRESENTATION BASED ON SPARSE LEARNING

受属性选择处理高维数据表现的高效性和低秩自表达方法在子空间聚类上成功运用的启发,提出一种基于稀疏学习的自表达属性选择算法。算法首先将每个属性用其他属性线性表示得到自表达系数矩阵;然后结合稀疏学习的理论(即整合L2,1-范数为稀疏正则化项惩罚目标函数)实现属性选择。在以分类准确率和方差作为评价指标下,相比其他算法,实验结果表明该算法可更高效地选择出重要属性,且显示出非常好的鲁棒性。

Inspired by the high efficiency of feature selection in dealing with high-dimensional data and the success application of low-rank self-representation in subspace clustering,we proposed a sparse learning-based self-represented feature selection algorithm. The algorithm first represents every feature in other feature linearity to obtain the self-representation coefficient matrix; then in combination with sparse learning theory( i. e. to integrate the L2,1-norm as a sparse regularisation punishment object function) it implements feature selection. With the evaluation indexes of classification accuracy and variance,and compared with the algorithms to be contrasted,experimental results indicated that the proposed algorithm could be more efficient in selecting important features and showed excellent robustness as well.