Heterogeneous feature selection with group structure mining in fuzzy decision systems for medical diagnosis

Abstract

In practical applications such as medical diagnosis and group decision making, the potential structural information contained in multi-dimensional features in the form of group domains plays an important role. However, most existing feature selection methods adopt transformed feature spaces for group structure analysis, which lack intrinsic semantic information interpretation. Meanwhile, fuzzy and uncertain heterogeneous data acquired from multiple devices increase the difficulty of task learning. Motivated by these two issues, this work devises a Heterogeneous Feature Selection method with Group Structure Mining in fuzzy decision systems (HFS-GSM), which follows the principle of one “strategy" and one “mechanism". Specifically, a feature group generation strategy based on fuzzy approximation Markov blanket is first designed for mining features with group structure, which introduces the concept of Markov blanket into the fuzzy rough set and utilizes the idea of approximation and fuzzy uncertainty measures. Then, a fuzzy dependency-based overlapping group elimination mechanism is proposed by attribution division, which avoids local redundancy while preserving global discriminative information. Furthermore, the effectiveness of HFS-GSM is verified in comparison with seven representative feature selection methods on publicly available medical datasets. Finally, medical diagnosis data provided by a hospital are obtained to demonstrate the reliability and utility of HFS-GSM in practical applications.