Algorithm 1.

Pseudo-code for ReliefF algorithm as implemented in ReBATE

Require: for each training instance a vector of feature values and the class value

1: n ← number of training instances

2: a ← number of attributes (i.e. features)

3: Parameter: k ← number of nearest hits ‘H’ and misses ‘M’

4:

5: # STAGE 1

6: pre-process dataset {≈ a · n time complexity}

7: # STAGE 2

8: pre-compute distance array {≈ 0.5 · a · n2 time complexity}

9: # STAGE 3

10: initialize all feature weights W[A] := 0.0

11: for i:=1 to n do

12:     # IDENTIFY NEIGHBORS

13:     for j:=1 to n do

14:         identify k nearest hits and k nearest misses (using distance array)

15:     end for

16:     # FEATURE WEIGHT UPDATE

17:     for all hits and misses do

18:         for A:= to a do

19:             W[A] := W[A] − diff (A,Ri,H)/(n · k) + diff (A,Ri,M)/(n · k)

20:         end for

21:     end for

22: end for

23: return the vector W of feature scores that estimate the quality of features