A group incremental approach for feature selection on hybrid data

Feature selection for dynamic data sets has been perceived as a very significant hot research problem in data mining. In practice, most real-world data usually are hybrid, which means both include categorical data and numerical data. For dynamic hybrid data, this paper first introduces a new neighborhood relation and information entropy based on neighborhood accordingly. Secondly, the single incremental mechanism and group incremental mechanism are analyzed and proofed to construct feature significance. On this basis, two incremental approaches to feature selection are developed for dealing with hybrid data. To better demonstrate the new algorithm, four common classifiers and twelve UCI data sets are introduced in the experiments. The experimental results further validate the feasibility of the incremental algorithms, and especially the efficiency of the group incremental algorithm.