IEEE Computational Intelligence Magazine - November 2022 - 63

TABLE IV The MAUC values of EMT-MC and EMT-MC-NonM on different datasets.
DATASET
EMT-MC
Balance
Contraceptive
Iris
Newthyroid
Wine-quality-red
Dermatology
Poker1
Segment
Led7digit
Penbased
win/tie/loss
denote that the result obtained by the
proposed method is significantly better,
significantly worse, and similar to that of
the baseline, respectively.
It can be observed from Tables III
and IV that no matter measured by
AveAcc or MAUC, the proposed
EMT-MC performs better than EMTMC-NonM.
The win/tie/loss statistics
also reveal that our algorithm wins or
ties on almost all of the datasets, and
only loses once. The reason that EMTMC
outperforms EMT-MC-NonM is
attributed to the fact that by introducing
the idea of multitasking learning,
useful information in " well-solved "
tasks can be transferred to " ill-solved "
tasks, which results in the improving
performance ofEMT-MC.
Moreover, it should be noted that
the suggested multitasking learning
0.9606 0.0001
0.7000 0.0010
0.9959 0.0001
0.9995 0.0001
0.8043 0.0007
0.9943 0.0001
0.8127 0.0006
0.9801 0.0001
0.9644 0.0001
0.9914 0.0001
EMT-MC-NONM
0.9604 0.0002 ()
0.7042 0.0009 ()
0.9820 0.0001 (+)
0.9862 0.0001 (+)
0.7249 0.0025 (+)
0.9946 0.0001 ()
0.7409 0.0112 (+)
0.9629 0.0001 (+)
0.9640 0.0001 ()
0.9845 0.0001 ()
5/5/0
idea notonlyenhancesthe qualityof
final classifiers but also accelerates the
convergence speed of EMT-MC. To
verify this fact, the convergence
curves of EMT-MC and EMT-MCNonM
on different datasets are plotted.
To be specific, the accuracy FEVs
(functionerror values)[50] isadopted
as the convergence metric which is
defined as the difference between the
objective function value of the found
solution and that of the global optimum
and is often used to evaluate the
convergence of an evolutionary multitasking
method [50]. The comparison
resultsare showninFig.2. Itcan be
found from the figure that compared
with EMT-MC-NonM, the proposed
EMT-MC has much better convergence,
which indicates the fact that by
introducing the multitasking learning,
the convergence speed of EMT-MC
improves a lot.
The comparison results in Tables III,
IV and Fig. 2 justify the effectiveness of
the suggested multitasking learning idea
in solving the decomposed-based multiclass
classification problem.
C. Comparison Between the Proposed
Algorithm and the Baselines
In the second part of experiments, the
proposed EMT-MC is compared with
four baselines on 10 datasets. Tables V
and VI present the experimental
results of different multiclass classification
algorithms in terms of AveAcc
and MAUC.
From the two tables, the following
observations can be obtained. Firstly, on
most of the datasets, two MOEAs
(MROCF and the proposed EMT-MC)
outperform the other three algorithms
(RUS, ELM and MC2ESVM) in terms of
AveAcc andMAUC.The reason is mainly
attributed to the following two facts. 1)
MROCF and EMT-MC adopt OVO
scheme which can obtain higher accuracy
than the algorithms with other schemes,
such as OVA (ELM, MC2ESVM), and
sampling technique (RUS). 2) These two
algorithms use TPR and FPR as two optimized
objectives, which can achieve good
performance on both balanced and imbalanced
datasets. Secondly, by further comparing
EMT-MC with MROCF, it can
be found that although both of them
adopt OVO scheme, EMT-MC achieves
higher AveAcc and MAUC values on
FIGURE 2 The convergence curves of EMT-MC and EMT-MC-NonM on different datasets.
NOVEMBER 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 63
IEEE Computational Intelligence Magazine - November 2022

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