关键词: 特征选择, 数据分类, 函数发现, 种群初始化, 层次邻域

Abstract: Abstract: Aiming at the shortcomings of general feature selection algorithms that only provide important features, but do not reveal the interpretable mapping relationship between features and data categories, an improved hierarchical distance gene expression programming feature selection classification algorithm is proposed. This method firstly uses the defined selection probability to make the head of the individual focus on the selection of function symbols during population initialization to increase the number of effective individuals in the initial population; secondly, define the layer neighborhood of individuals to make the population individuals select symbols in its layer neighborhood when mutating, change the blind and undirected problem of hierarchical variation; finally, combine the individual’s dimensionality reduction rate and classification accuracy as the individual’s fitness value, change the evolutionary model of the population’s single optimization goal, and balance the two relationship between people. Perform 5-Fold or 10-Fold verification on 7 data sets. The proposed algorithm gives a concise and clear function mapping relationship between data features and their categories. The mapping function obtained by the algorithm is classified, compared with other algorithms, the dimensional reduction rate of the Hepatitis, WPBC, Sonar, and WDBC data sets are imporoved by 9.47%, 1.34%, 0.17%, and 1.34% respectively. The average classification accuracy rate of Hapatitis, Ionosphere, Musk1, WPBC, Heart-Statlog, WDBC data sets are improved by 1.04%, 1.04%, 2.31%, 2.89%, 1.7%, 1.41% respectively. Experimental results verify the feasibility, effectiveness and superiority of the proposed algorithm in feature selection and classification.

Key words: feature selection, data classification, function discovery, population initialization, layer neighborhood