Abstract:
To address the issue of feature selection for relative poverty identification, an orthogonal selection method based on pairwise-sample comparison was proposed. Paired sample sets of “relative poverty” and “non-relative poverty” were collected by means of pairwise by means of pairwise comparison. Then, a new feature subset evaluation function was designed based on the idea of pulling similar samples closer and pushing dissimilar samples further apart. Finally, orthogonal experimental design was employed to select features. To validate the effectiveness of the method, 356 registered poor households and 212 non-registered poor households from the Dabie Mountain area were considered as research subjects. Four sets of paired sample sets were randomly constructed to screen four groups of key features,and seven classifiers including logistic regression, decision tree, support vector machine, deep neural network, random forest, Boosting, and naive Bayes were tested for performance evaluation. The results indicate that,with the exception of the decision tree, accuracy, sensitivity, specificity, and AUC values exceeding 90% are achieved by the other six classifiers across all four sets of key features. Minimal variation is observed in the identification performance of features selected from different sample sets, and comparable performance to that of the full feature set is attained by all four sets of key features.The proposed method is characterized by its simple principle and operational convenience, making it suitable for scenarios where relative poverty classification standards are lacking or difficult to establish, thereby enabling effective screening of identification features.