Tolerance-based and fuzzy-rough feature selection

One of the main obstacles facing the application of computational intelligence technologies in pattern recognition (and indeed in many other tasks) is that of dataset dimensionality. To enable pattern classifiers to be effective, a dimensionality minimization step is usually carried out beforehand. Rough set theory has been successfully applied for this as it requires only the supplied data and no other information; most other methods require supplementary knowledge. However, the main limitation of traditional rough set-based selection in the literature is the restrictive requirement that all data is discrete; it is not possible to consider real-valued or noisy data. This has been tackled previously via the use of discretization methods, but may result in information loss. This paper investigates two approaches based on rough set extensions, namely fuzzy-rough and tolerance rough sets, that address these problems and retain dataset semantics. The methods are compared experimentally and utilized for the task of forensic glass fragment identification.