Q2. How many measures were selected for each of the fuzzy-rough measures?
For each of the fuzzy-rough measures introduced, the authors ran QuickReduct once with α = 1, and a second time with a fixed α < 1; in particular, a value of α = 0.95 was deemed a suitable overall choice for most measures, except for g, which requires a much higher threshold, and for which α = 0.9999 was selected.
Q3. What is the way to interpret the results?
When interpreting these results, one should always keep in mind the trade-off between accuracy (RMSE) and attribute subset size: a higher accuracy (lower RSME) is of course desirable, but so is a smaller subset size, i.e., the less conditional attributes there are in the reduced data set, the stronger its generalization capacity.
Q4. What is the general definition of fuzzy decision reducts?
In the general definition that the authors propose, the authors require an increasing [0, 1]-valued measure, so as to guarantee that the larger an attribute subset, the higher its degree of fuzzy decision reducthood (monotonicity), which is in analogy to other approaches to define a degree of approximating decision classes [43, 44].
Q5. What are the evaluation measures in the previous subsections?
As the authors have shown, the evaluation measures γ, γ′, δ, δ′, f and g introduced in the previous subsections all give rise to corresponding fuzzy decision reducts.
Q6. How much accuracy can be achieved with fuzzy -decision reducts?
as seen in Figure 2b), if a 1% accuracy drop is permissible, fuzzy γ-decision reducts manage to reduce the subset size by over 40%, while with g a reduction of the data set by more than 63% is possible.
Q7. What is the effect of the heuristic on the and f subset?
This affects QuickReduct’s operation adversely; when all of the considered subsets in a given iteration evaluate to 0, the heuristic is forced to select one without any information about its true merit.
Q8. What is the default distance weighting for the K-nearest neighbour classifier?
In their experiments, the authors have used the very simple K-nearest neighbour classifier [1], implemented in Weka [53] as IBk, with default parameters (K = 1, no distance weighting).
Q9. What is the difference between the two experiments?
Their experiments clearly endorse the benefit of using fuzzy decision reducts, showing a greater flexibility and better potential to produce good-sized, highquality attribute subsets than the crisp decision reducts that have been used so far in fuzzy-rough data analysis.