feature selection for classification：分类特征选择

dure. The fact that an evaluation function must be monotonic to be applicable to these methods, prevents the use of mon evaluation functions. This problem is partially solved by relaxing the monotonicity criterion and introducing the approximate monotonicity concept [ 161. 3.2.2. Hand-run of CorrAL Dataset (see Figure 3) The authors use Mahalanobis distance measure as the evaluation function. The algorithm needs input of the required number of features (M) and it attempts to find out the best subset of N 2M features to B&B(D, S, M) if (curd(S) <> M) { /* subset generation 7 j=O For all features f in S { Sj = S - f/* remove one feature at a time */ if is legitimate) */ if ZsBetter( Sj , T) T = Sj /* recursion */ B&WSj, W} j++ 1 return T }) Fig. 3. Branch and bound.