IEEE Computational Intelligence Magazine - May 2022 - 77

effectiveness of knowledge transfer in
EMTAUC is confirmed.
In addition, the average convergence
trends of four compared methods, i.e.,
EMTAUC-MFEAII, EMTAUC-SBGA,
EMTAUC-EMEA, and Single-task GA,
on usps, a9a, ijcnn1, and poker datasets
are shown in Fig. 4 to illustrate the effectiveness
of our proposal further. Due to
the space limitation, the average convergence
trends of four compared methods
on all datasets are shown in Fig. S2 (see
Supplementary Material). In these figures,
the x-axis represents the computational
cost, and the y-axis represents the average
objective value on a log scale. As can be
seen from Fig. 4, EMTAUC-MFEAII,
EMTAUC-SBGA, and EMTAUC-EMEA
show excellent performance in the average
convergence trend compared with
Single-task GA in most cases, which
illustrates that knowledge transfer across
tasks can effectively accelerate the convergence
of expensive tasks. In addition,
EMTAUC-MFEA-II significantly outperforms
the other multitasking methods,
especially on the poker dataset. As shown
in Fig. 3, the intertask Spearman rank
correlation is not high on the poker dataset,
so harmful knowledge across tasks
may prevail in the complex multitasking
AUC environment. MFEAII can theoretically
minimize the trend of negative
transfer across tasks, while other methods
do not explicitly consider preventing the
process, such as SBGA and EMEA.
Therefore, EMTAUC-MFEA-II shows
higher performance on most datasets.
The experiments are designed to compare
EMTAUC with online learning
methods, as shown in Table IV. The
symbols " − " , " ≈ " and " + " imply that the
corresponding compared method is significantly
worse, similar to, and better
than EMTAUC on the Wilcoxon ranksum
test with a 95% confidence level,
respectively. The AUC of EMTAUC
outperforms CW, PA, and Perceptron in
9 of 10 cases and loses once to them on
the vowel dataset. EMTAUC outperforms
CPA in 8 out of 10 cases and
loses two times on vowel and poker
datasets. OAMinf exceeds EMTAUC in
2 of 10 cases, matches two times, and
loses six times to EMTAUC. EMTAUC
MAY 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 77
TABLE IV The comparison of EMTAUC against gradient-based methods in terms of AUC.
DATASET
diabetes
OAMseq
fourclass
german
splice
usps
a9a
sonar
svmguide1
svmguide3
segment
ijcnn1
satimage
vowel
poker
w/t/l
0.826(0.037) −
0.831(0.025) .
0.775(0.041) −
0.859(0.019) −
0.931(0.016) −
0.842(0.017) −
0.850(0.042) −
0.988(0.003) .
0.760(0.035) −
0.956(0.013) −
0.921(0.008) −
0.919(0.014) −
0.931(0.046) +
0.592(0.060) +
2/2/10
OAMgra
0.826(0.034)−
0.829(0.025) .
0.772(0.036)−
0.886(0.017)+
0.935(0.012)−
0.857(0.017)−
0.849(0.043)−
0.988(0.002) .
0.755(0.034)−
0.955(0.014)−
0.910(0.009)−
0.911(0.017)−
0.928(0.046)+
0.586(0.062)+
3/2/10
OAMinf
−
0.831(0.020) .
0.776(0.034)−
−
−
−
0.849(0.043)−
0.989(0.002) .
0.768(0.035)−
0.956(0.013)−
0.928(0.015) .
0.921(0.013)−
0.931(0.041)+
0.594(0.067)+
2/2/6
CPA
−
0.812(0.020)−
0.701(0.033)−
−
−
−
0.827(0.052)−
0.885(0.008)−
0.707(0.037)−
0.886(0.021)−
0.910(0.011)−
0.828(0.024)−
0.923(0.032)+
0.524(0.067)+
2/0/8
CW
−
0.758(0.032)−
0.757(0.025)−
−
−
−
0.793(0.059)−
0.922(0.019)−
0.723(0.036)−
0.896(0.021)−
0.829(0.021)−
0.619(0.024)−
0.870(0.063)+
0.457(0.058)−
1/0/9
PA
−
0.668(0.0168)−
0.701(0.033)−
−
−
−
0.790(0.057)−
0.799(0.219)−
0.696(0.038)−
0.852(0.024)−
0.531(0.074)−
0.646(0.024)
0.863(0.063)+
0.506(0.028)−
1/0/9
PERCEPTRON
−
0.690(0.165)−
0.701(0.039)−
−
−
−
0.780(0.060)−
0.883(0.140)−
0.696(0.038)−
0.852(0.024)−
0.647(0.088)−
0.605(0.025)−
0.859(0.055)+
0.502(0.031)−
1/0/9
EMTAUC
0.833(0.002)
0.832(0.0003)
0.796(0.001)
0.878(0.001)
0.965(0.0007)
0.896(0.0001)
0.856(0.010)
0.989(0.0001)
0.795(0.004)
0.987(0.0008)
0.927(0.006)
0.999(0.00003)
0.806(0.001)
0.517(0.0001)
−
IEEE Computational Intelligence Magazine - May 2022

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