Graphics Reference
In-Depth Information
Table 4.6
Average ranks for the Lazy Learning methods
1-NN
3-NN
LBR
LWL
Avg.
Ranks
IM
5
11
5
8
7.25
7
EC
9.5
13
9
8
9.88
12
KNNI
2.5
5.5
9
8
6.25
4
WKNNI
4
5.5
9
8
6.63
5
KMI
12
5.5
9
2.5
7.25
8
FKMI
6
1.5
9
2.5
4.75
3
SVMI
9.5
9
3
8
7.38
9
EM
11
5.5
9
2.5
7.00
6
SVDI
13
12
1
12
9.50
11
BPCA
14
14
13
13
13.50
14
LLSI
7.5
5.5
9
8
7.50
10
MC
7.5
1.5
3
2.5
3.63
1
CMC
1
5.5
3
8
4.38
2
DNI
2.5
10
14
14
10.13
13
the third best, FKMI. No other family of classifiers present this gap in the rankings.
Therefore, in this family of classification methods we could, with some confidence,
establish the EC method as the best choice. The DNI and IM methods are among
the worst. This means that for the black-boxes modelling methods the use of some
kind of MV treatment is mandatory, whereas the EC method is the most suitable one.
As with the RIL methods, the BPCA method is the worst choice, with the highest
ranking.
Finally the results for the last LL group are presented in Table
4.6
.FortheLL
models, the MC method is the best with the lowest average ranking. The CMC
method, which is relatively similar to MC, also obtains a low rank very close to
MC's. Only the FKMI method obtains a low enough rank to be compared with
the MC and CMC methods. The rest of the imputation methods are far from these
lowest ranks with almost two points of difference in the ranking. Again, the DNI and
IM methods obtain high rankings. The DNI method is one of the worst, with only
the BPCA method performing worse. As with the black-boxes modelling models,
the imputation methods produce a significant improvement in the accuracy of these
classification methods.
4.6.3 Interesting Comments
In this last Section we have carried out an experimental comparison among the impu-
tation methods presented in this chapter. We have tried to obtain the best imputation
choice by means of non-parametric statistical testing. The results obtained concur
with previous studies:
