Are More Features Better? A Response to Attributes Reduction Using Fuzzy Rough Sets

A recent TRANSACTIONS ON FUZZY SYSTEMS paper proposing a new fuzzy-rough feature selector (FRFS) has claimed that the more attributes remain in datasets, the better the approximations and hence resulting models. [Tsang et al., IEEE Trans. Fuzzy Syst., vol. 16, no. 5, pp. 1130–1141]. This claim has been used as a primary criticism of the original FRFS method [Jensen and Shen, IEEE Trans. Fuzzy Syst., vol. 15, no. 1, pp. 73–89, Feb. 2007]. Although, in certain applications, it may be necessary to consider as many features as possible, the claim is contrary to the motivation behind feature selection concerning the curse of dimensionality, the presence of redundant and irrelevant features, and the large amount of literature documenting observed improvements in modeling techniques following data reduction. This letter discusses this issue, as well as two other issues raised by Tsang et al. [IEEE Trans. Fuzzy Syst., vol. 16, no. 5, pp. 1130–1141, Oct. 2008] regarding the original algorithm.