Graphics Reference
In-Depth Information
the number of wins and ties obtained by the method in the row. The maximum value
for each column is highlighted in bold.
Observing Tables.
8.8
and
8.9
, we can point out the best performing PS methods:
•
In condensation incremental approaches, all methods are very similar in behav-
ior, except PSC, which obtains the worst results. FCNN could be highlighted in
accuracy/kappa performance and MCNN with respect to reduction rate with a low
decrease in efficacy.
•
Two methods can be emphasized in from the condensation decremental family:
RNN and MSS. RNN obtains good reduction rates and accuracy/kappa perfor-
mances, whereas MSS also offers good performance. RNN has the drawback of
being quite slow.
•
In general, the best condensation methods in terms of efficacy are the decremental
ones, but their main drawback is that they require more computation time. POP
and MSS methods are the best performing in terms of accuracy/kappa, although
the reduction rates are low, especially those achieved by POP. However, no con-
densation method is more accurate than 1NN.
•
With respect to edition decremental approaches, few differences can be observed.
ENN, RNGE and NCNEdit obtain the best results in accuracy/kappa and MENN
and ENNTh offers a good tradeoff considering the reduction rate. Multiedit and
ENRBF are not on par with their competitors and they are below 1NN in terms of
accuracy.
•
AllKNN and MoCS, in edition batch approaches, achieve similar results to the
methods belonging to the decremental family. AllKNN achieves better reduction
rates.
•
Within the hybrid decremental family, three methods deserve mention: DROP3,
CPruner and NRMCS. The latter is the best, but curiously, its time complexity
rapidly increases in the presence of larger data sets and it cannot tackle medium size
data sets. DROP3 is more accurate than CPruner, which achieves higher reduction
rates.
•
Considering the hybrid mixed+wrapper methods, SSMA and CHC techniques
achieve the best results.
•
Remarkable methods belonging to the hybrid family are DROP3, CPruner, HMNEI,
CCIS, SSMA, CHC and RMHC. Wrapper based approaches are slower.
•
The best global methods in terms of accuracy or kappa are MoCS, RNGE and
HMNEI.
•
The best global methods considering the tradeoff reduction-accuracy/kappa are
RMHC, RNN, CHC, Explore and SSMA.
8.6.2 Analysis and Empirical Results on Medium Size Data Sets
This section presents the study and analysis of medium size data sets and the best
PS methods per family, which are those highlighted in bold in Table
8.8
. The goal
