基于?2,1范数的在线流特征选择算法

Online Streaming Feature Selection Algorithm Regularized by ?2,1-norm

高维特征数据包含大量的无关信息和冗余信息,这些信息可能会极大降低学习算法的效率。对于加速机器学习算法,提升学习模型泛化能力和避免维数灾难的影响,特征选择算法在很多应用场景下扮演重要角色。在数据特征空间未知,动态变化的场景下,传统的基于静态特征空间场景的特征选择算法因效率低而不适用。为解决特征空间动态未知的流特征场景下的特征选择问题,提出基于l2,1范数的在线流特征选择算法。利用l2,1范数的行稀疏性质和噪声不敏感的特性实现特征选择模型的构建。实验表明,在多个高维数据集上,新提出的流特征选择算法相比较其他的流特征选择算法具有较高的分类识别率和稳定性。

High dimensional streaming feature data contain a mass of irrelevant and redundant information,which may greatly reduce the efficiency of learning algorithms. Feature selection algorithms play an important role in many application scenarios for speeding up machine learning algorithms,improving the generalization ability of learning models and avoiding the curse of dimensionality. In the scene where the feature space is unknown and dynamic,the traditional feature selection algorithm based on the static feature space is not suitable for low efficiency. In order to solve streaming feature selection problem that feature space is dynamic and unknown,the paper proposes the online streaming feature selection regularized by l2,1-norm. The paper constructes the feature selection model using the sparse property of the l2,1-norm and the insensitivity of the noise. Experimental results demonstrate that,compared with other streaming feature selection algorithms,the proposed feature selection algorithm has higher recognition performance and stability in multiple high-dimensional datasets.