Online Hierarchical Feature Selection Algorithms With Streaming Features

As a significant data pre-processing method,feature selection effectively reduces the training time and improves the learning accuracy of learning models.However,with the rapid speeding up of data generation and collection in the era of big data,traditional feature selection algorithms are confronted with serious challenges:(1)the existence of high dimensionality in feature space is usually accompanied by unknown and evolution;(2)the categories are not often independent of each other,and there usually exist complex hierarchical relationships.These result in a problem that traditional feature selection algorithms are not effective and even fail to work in real application environments.In this dissertation,we address the problem of fully exploiting and utilizing the hierarchical structure of classes for feature selection with a dynamic feature space,and conduct research on online hierarchical feature selection methods with streaming features.Specifically,the main research works of this paper are as follows.(1)Online hierarchical feature selection based on neighborhood rough sets with streaming features.Traditional online streaming feature selection algorithm ignores the hierarchical structure relationship among classes of samples.To address this problem,a new neighborhood rough model for hierarchical structured data is defined by using the sibling strategy among nodes in the category hierarchy.Secondly,an online streaming feature selection framework for hierarchical classification is constructed including feature online importance selection and online redundancy update based on neighborhood dependency.The corresponding online streaming hierarchical feature selection algorithm is proposed.Experiments are conducted to verify the effectiveness of the algorithm.(2)Online hierarchical feature selection based on kernelized fuzzy rough sets with streaming features.The sibling strategiy among nodes in the category hierarchy defines a new kernelized fuzzy rough model for hierarchically structured data to address the problem that traditional online streaming feature selection algorithm ignores the hierarchical structure relationships between among categories,to further optimize the time performance of online operations and to efficiently measure the fuzzy relationships between among samples.Secondly,on the basis of the streaming feature selection framework in(1),a new online streaming hierarchical feature selection algorithm is proposed by redefining screening strategy of feature online importance selection and online redundancy update phase using kernel fuzzy dependencies.The experimental results show a further improvement in the performance of the algorithm.