IEEE Computational Intelligence Magazine - November 2022 - 68

TABLE XVI The weighted macro-F1 scores of different algorithms on the datasets with more complex characteristics.
DATASET
EMT-MC
Arrhythmia
Movement libras
Isolet5
letter-recogmition
Pendigits
Poker2
Shuttle
Urban
win/tie/loss
0.4984 0.0051
0.6605 0.0025
0.7014 0.0018
0.5095 0.0012
0.7318 0.0010
0.3715 0.0041
0.8625 0.0007
0.3757 0.0065
RUS
0.2918 0.0038(+)
0.2526 0.0032(+)
0.4801 0.0010(+)
0.3083 0.0001()
0.7911 0.0001(-)
0.2021 0.0003(+)
0.8359 0.0002(+)
0.2801 0.0074(+)
7/0/1
ELM
0.4542 0.0032(+)
0.1511 0.0016(+)
0.0141 0.0001(+)
0.0307 0.0001(+)
0.3109 0.0004(+)
0.0325 0.0015(+)
0.7341 0.0002(+)
0.2204 0.0033(+)
8/0/0
MC2ESVM
0.4285 0.0012(+)
0.3028 0.0018(+)
0.5049 0.0003(+)
0.4192 0.0001(+)
0.6768 0.0001(+)
0.2011 0.0001(+)
0.7871 0.0001(+)
0.2428 0.0011(+)
8/0/0
MROCF
0.4455 0.0038(+)
0.6461 0.0031(+)
0.7287 0.0007(-)
0.7316 0.0001(-)
0.9124 0.0003(-)
0.2704 0.0139(+)
0.8258 0.0007(+)
0.3628 0.0029(+)
5/0/3
complex dataset has more sub-problems
(tasks). Based on the suggested strategies
in EMT-MC, there are much more
assisting operations among different tasks,
which may incur some negative transfers
among tasks. Thus, the superiority of
EMT-MC is not so obvious. Unlike
EMT-MC the sub-problems in
MROCF run independently, which
means more numbers of sub-problems
do not affectMROCF very much.
Moreover, Table XVII gives the
time comparison of different algorithms
on the complex datasets. The
comparison results show that the running
time of five algorithms is similar
to that on 10 simple datasets (see
Table VII). Specifically, RUS algorithm
has the fastest running time on
most of the datasets. MROCF and
EMT-MC have the longest running
time, since both of them adopt the
OVO decomposition scheme which
has to tackle mðm 1Þ=2sub-problems.
Furthermore, EMT-MC has
more running time than MROCF. It
is because that each sub-problem in
MROCF is solved independently,
and there is no assisting operations
between different sub-problems,
which saves much computational time.
Finally, for further analyzing performance
of the proposed EMT-MC, a
series of additional experiments are also
conducted from different aspects. These
experiments mainly include two parts.
1) The comparison between the proposed
method and classical multiclass
classification algorithms. 2) Parameter
analysis on the proposed EMT-MC.
For better analyzing the performance
of the proposed EMT-MC, a series of
additional experiments have been conducted
from different aspects (presented
in the supplementary materials available
online).
V. Conclusion
In this article, an evolutionary multitasking
method (EMT-MC) was proposed
for multiclass classification. InEMT-MC,
the OVO strategy was firstly utilized to
decompose the original multiclass classification
problem into several binary classification
sub-problems. Then, each binary
classification sub-problem was viewed as
TABLE XVII The time comparison (s) of different algorithms on the datasets with more
complex characteristics.
DATASET
Arrhythmia
Movement libras
Isolet5
letter-recogmition
Pendigits
Poker2
Shuttle
Urban
EMT-MC
38530
35975
682416
714769
4338
3498
4134
1875
RUS
1.1
0.5
89
59
9
15
29
1.1
ELM
176
123
884
5075
1205
216
705
34
MC2ESVM MROCF
1.7
0.6
5.2
1238
2183
332
189
178
388
0.9
14280
15888
955
935
1060
477
a task and evolved with an independent
population. During the evolution, for
each " ill-solved " task T, the idea ofevolutionary
multitasking learning was introduced
to construct its corresponding
" assisting " task set and help task T to
achieve the classifiers with higher accuracy.
The experimental results on different
multiclass classification datasets
justified the competitiveness of the proposed
method in comparison with the
state-of-the-art algorithms.
The results of this work suggest that
the evolutionary multitasking learning is
a promising idea for solving the decomposition-based
multiclass classification
problem. Along with this research line,
there also exist several interesting issues to
be further investigated. For example, in
this paper, the multiclass classification
problem was solved by using the strategy
ofOVO which decomposes an m-classes
classification problem into mðm 1Þ=2
binary classification sub-problems. However,
if the value of m is large, it means
there are too many sub-problems. How
to design an efficient evolutionary multitasking
algorithm under this situation is
one of our future plans. In addition, it
is well-known that besides the OVO
strategy, OVA is another famous
decomposition strategy for multiclass
classification. However, it may create
many imbalanced (even extremely
imbalanced) classification sub-problems.
How to develop an effective evolutionary
multitasking algorithm with
OVAstrategy isalsoaninteresting
topic to be further tackled. Lastly, as
mentioned in Section IV.E, the
F1-score is another useful metric that
can measure the quality ofthe obtained
68 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2022
IEEE Computational Intelligence Magazine - November 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2022
