T-FWH: A Hybrid Feature Selection Method Combining Multi-Criteria Decision Making for Imbalanced Data

In the presence of imbalanced data problem, existing feature selection methods focus on retaining features that are useful for majority class sample recognition, which leads to poor performance of recognizing minority class samples. However, in fields such as healthcare and finance, the identification of minority class samples is often of paramount importance. To this end, we design a novel hybrid feature selection method combining multi-criteria decision making, which is named T-FWH. On the one hand, the T-FWH integrates filter and wrapper techniques to construct a two-stage feature selection algorithm, which not only improves the efficiency of existing feature selection methods but also ensures the quality of feature selection; on the other hand, we comprehensively evaluate the candidate feature subsets in combination with multi-criteria decision-making method and determine the optimal feature subset, thus solving the problem of feature loss of minority class samples due to imbalanced data. The experiments on a series of high-dimensional datasets demonstrate that the proposed method can effectively reduce the redundant features in the data, and the use of the selected features can improve the classification accuracy of the model for minority class samples.