A fuzzy rough set approach to hierarchical feature selection based on Hausdorff distance

摘要

With increases in feature dimensions and the emergence of hierarchical class structures, hierarchical feature selection has become an important data preprocessing step in machine learning. A variety of effective feature selection methods based on granular computing and hierarchical information have been proposed. The fuzzy rough set method is an effective granular computing method for dealing with uncertainty. However, it is time-consuming because the distance calculations are only based on single samples. In this paper, we propose a fuzzy rough set approach using the Hausdorff distance of the sample set for hierarchical feature selection. This integrates the benefits of sample granularity and class hierarchical granularity. Firstly, the general feature selection task is decomposed into coarse-grained and fine-grained tasks according to the hierarchical structure of the data’s semantic labels. This allows a large and difficult classification task to be divided into several small and controllable subtasks. Then, the Hausdorff distance-based fuzzy rough set method is used to select the best feature subset in each coarse- and fine-grained subtask. Unlike single-sample-based distance calculation, Hausdorff distance calculation uses a sample set of different classes. The new model greatly reduces the computational complexity of classification. Finally, we use the top-down support vector machine classifier to experimentally verify the effectiveness of the proposed methods on five hierarchical datasets. Compared with five existing feature selection algorithms in terms of three evaluation metrics, the proposed method provides the highest average accuracy and much lower running time. In particular, on the F194 dataset, our method takes the least time to improve the FH indicator by 2% compared with that of the second-best algorithm.