Professional Documents
Culture Documents
LaxmiLydia2022 Chapter OptimalBoostingLabelWeightingE
LaxmiLydia2022 Chapter OptimalBoostingLabelWeightingE
LaxmiLydia2022 Chapter OptimalBoostingLabelWeightingE
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1
A. Khamparia et al. (eds.), Biomedical Data Analysis and Processing Using Explainable
(XAI) and Responsive Artificial Intelligence (RAI), Intelligent Systems Reference
Library 222, https://doi.org/10.1007/978-981-19-1476-8_1
2 E. Laxmi Lydia et al.
1.1 Introduction
XAI might be a potential avenue for guiding clinical interventions and fundamentals
of neural circuit manipulation.
Jesmin et al. [9] designed an AI-driven cloud-based self-stress recognition method
that takes physiological signals like peripheral capillary oxygen saturation, heart rate
variability, galvanic skin response to describe the stress levels. The sensor nodes are
embedded with wearable devices and collect physiological signals and thus recognize
stress level of a person. Nahian et al. [10] present an emotion-aware fall monitoring
architecture with big data analytics, IoT, and AI technique that would handle senti-
ment analysis of the elderly people, prediction regrading real-time fall monitoring,
and health conditions. In emergency settings, the presented architecture alerts about
a case of an emergency to the predetermined caregivers. A mobile clinic or smart
ambulance would reach the older adult locations at minimal time.
Deshpande and Rao [11] focus on employing NLP method on Twitter feeds to
conduct sentiment analysis. The individual tweet is categorized as negative or neutral,
according to curated word list to identify the tendency of depression. In the procedure
of class predictions, SVM and NB classification were employed. Ćosić et al. [12]
analysed the capacity to use GPS data, passively acquired from smartphones over
seven days, to identify PTSD diagnosis status amongst a cohort (N = 185) of higher-
risk, formerly traumatized women. With the maximum distance travelled and daily
time spent away from home as a source for feature engineering model, the result
suggests that diagnosis group status could be forecasted with higher efficiency.
In Li et al. [13], projected a DL-based integrated psychological activity monitoring
system (DLIPAMS) for predicting the depression and anxiety of an individual. For
predicting mental disorders, the nature of ML and AI methods is extensively used.
While executed in real time, this system could assist society by serving as moni-
toring tool for individuals with immoral behaviours. Pabreja et al. [14] analyses
stress amongst students of some educational institutions in India. The information
has been gathered from 650 respondents with Likert scale of 5. With the applica-
tions of distinct data visualization methods and RF regression algorithms, 15 main
contributing factors from a list of 25 characteristics have been recognized and the
predictions of stress level have been performed by an R-squared value of 0.8042.
This article designs an optimal boosting label weighting extreme learning machine
for mental disorder prediction and classification (OBWELM-MDC) technique. The
OBWELM-MDC technique involves the design of boosting label weighted extreme
learning machine (BWELM) model for the prediction process. In addition, the
BWELM model can be derived by the incorporation of the label weighted extreme
learning machine (LW-ELM) with boosted ensemble learning model. Furthermore,
the parameter tuning of the BWELM model takes place by the use of chaotic star-
ling particle swarm optimization (CSPSO), where the inertia weight and acceler-
ation coefficient of the PSO algorithm are modified via logistic chaotic map. The
experimental result analysis of the OBWELM-MDC technique takes place using the
benchmark dataset, and the results are investigated under several measures.
4 E. Laxmi Lydia et al.
This article has designed an effective OBWELM-MDC technique for mental disorder
prediction and classification. The OBWELM-MDC technique follows a three-
stage process, namely pre-processing, BWELM-based prediction, and CSPSO-based
parameter tuning. In addition, the parameter tuning of the BWELM model takes place
by the use of CSPSO, where the inertia weight and acceleration coefficient of the PSO
algorithm are modified via logistic chaotic map. Figure 1.1 showcases the process
of XAI proposed method.
In an initial phase, the data pre-processed follows three states as data normalization,
class labelling, and format conversion. At present, the medical data in several formats
(i.e. csv) was changed to compatible. arff format. Next, data normalizing method
was min–max procedure. At this point, the maximal and minimal values from the
data were achieved and the values were normalized effectually. The purpose is to
normalize the input values as to range of zero and one and disseminate other values
to planned range. The normalization technique was gained by employed in Eq. (1.1):
x − xmin
Min − Max.Norm = (1.1)
xmax − xmin
Finally, the class labelling technique was applied that the data samples under the
dataset are assigned for appropriating class labels like zero and one.
At this stage, the pre-processed data is fed as input to the BWELM model for mental
disorder detection and classification. WELM method is considered as a cost-sensitive
learning version of ELM which is an efficient method to deal with imbalanced data
[15]. Like CS-SVM, the primary concept of WELM is to allocate distinct penalties
for several classes, whereas the minority class has a large penalty factor C. WELM
focuses on the training error of the minority instance, which makes a classification
hyperplane emerge in an impartial location. A weighted matrix W is utilized for
regulating the parameter C for distinct instances as
1
N
1
minimalize L pELM = β2 + C W ξi 2 (1.2)
2 2 i=1
1
WELM1 : Wii = ,
#(t )
i
0.618
#(ti )
if #(ti ) > AVG(ti )
WELM2 : Wii = (1.3)
1
#(ti )
if #(ti ) ≤ AVG(ti )
Let Wii , #(ti ), AUG (ti ), and 0.618 be the weight of ith training sample, the
amount of instance belongs to the class ti , and the average amount of instances over
each class. In comparison with WELM2, WELM1 is popular and practical [16].
6 E. Laxmi Lydia et al.
I −1
HT + WHHT WT, when N ≤ L
β= I C −1 T (1.4)
C
+ HWHT H W T, when N > L
whereas D1l (xi ) represents the primary label weight of the xi instance on the lth label,
N N
and Z 1l = |til | signifies the standardization factor D1l (xi ) = 1. In tth iteration
i=1 i=1
in the AdaBoost training process, D1l (xi ) and Z 1l are changed to be Dtl (xi ) and Z tl ,
correspondingly. Specifically, the first label weight of each instance on all the labels
has been scaled without the proportion variation as follows:
N
εrl = Dtl (xi ) (1.7)
i=1(xi )l =Yil
While N represents the amount of multi-label training samples, (xi )l and Yil
indicate the actual and predicted label status (inactive or active) of the sample xi on
the lth label, correspondingly. As per the concept of AdaBoost, we have the label
weight update parameter as follows:
N
i=1(xi )l =Yil Dt (x i )
l
1
αrl = ln N (1.8)
i=1(xi )l =Y l Dt (x i )
2 l
i
in which αrl represents the label weight update parameter related to the lth label.
Generally, αtl represents a positive number for every classifier, and it could precisely
1 Optimal Boosting Label Weighting Extreme Learning Machine … 7
identify further instances when compared to imprecise classification. Then, the new
label weight of the (t + 1)th iteration is upgraded by
Dtl (xi ) exp(−αtl I t xi )l , Yil
l
Dt+1 (xi ) = (1.9)
Z rl
N
and Z rl = Dl (xi ) continue to perform as a standardization factor to ensure
t
i=1
N
Dl (xi ) = 1. That implies the classifier could focus gradually on complex
t+1
i=1
combination of class instance and label. From other perspectives, the algorithm could
adoptively examine the label weight distribution without evaluating the likelihood
density.
Lastly, the voting weight of every base classifier in the ensemble method could
be estimated as
⎛ |Y | ⎞
1 1−1 − N
i=1(xi ) =Yi t
1 l D 1
(x i )
Wt = ln⎝ |Y | ⎠ (1.11)
2 11 = N
l Dt (x i )
1
i=1(xi ) =Yi
1
as well as the active label of the future unseen instance is determined by weighted
voting.
For optimally adjusting the parameters of the BWELM model, the CSPSO algorithm
is applied. PSO is the swarm intelligence algorithm stimulated by social behaviours
of bird flocking. In PSO, the population size of the particle is as follows as well as
the dimensions of the searching space are represented as D. All the particles i have
a location vector xi = {xi1 , xi2 , . . . , xi D }, a velocity vector νi = {νi1 , νi2 , . . . , νi D },
and an individual optimal location in history pi = { pi1 , pi2 , . . . , pi D }. The optimal
location found
by each particle
in the swarm named the global best is signified
as pg = pg1 , pg2 , . . . , pg D . In the iteration t, the novel upgraded location and
velocity of particle i are defined as follows:
8 E. Laxmi Lydia et al.
νi j (t + 1) = w ∗ νi j (t) + c1 ∗ r1 ∗ pi j (t) − xi j (t) + c2 ∗ r2 ∗ p+ j (t) − xi j (t) ,
(1.12)
whereas w represents the inertia weight, c1 and c2 denote the acceleration coefficient
controlling the step size as well as r1 and r2 indicate independently created arbitrary
numbers distributed uniformly amongst zero and one. The two end condition is
utilized. One is when a tolerance level was attained, and another one is when a
present maximal amount of iterations is obtained.
An extensive nonlinear phenomenon naturally has the features of ergodicity, sensi-
tivity, randomness to primary condition, etc. Because of randomness and ergodicity
characteristics, chaotic motion could traverse each state to a specific extent based
on their own laws without repetition. Hence, utilizing chaos variables for optimally
searching, we would undoubtedly have more benefits when compared to random
search. Chaos ergodicity characteristics are utilized for optimizing the search and
preventing falling into the local minimal; thus, chaos optimization search method-
ology has become a new optimization approach. Chaotic sequences produced by
distinct mappings are employed, i.e. tent map, sine map, logistic map, sinusoidal
map, and singer map. In this work, various chaotic maps are tried and the optimal
one was selected for combining the PSO approach. Based on the preliminary research,
logistic map has attained better outcomes.
pch = xi ∗ θ (1.15)
Individual factor social factor and velocity of the previous iteration are employed
for updating velocity of the particles. Three variables, r1 , and r2 , control the rela-
tive contribution of the three elements. The inertia weight w controls the influ-
ences of the velocity of the preceding iteration. The cognitive coefficient c1 and c2
balances the contribution of the social and individual factors. Moreover, the accel-
eration coefficient r1 and r2 controls the proportion maintained in the social and
individual factors. All the parameters play a significant role in the velocity updating
and subsequently affect the particle location updating. Rather than utilizing fixed
value, logistic mapping is utilized for generating the acceleration coefficients r1 and
r2 in every generation. Also, the inertia weight is altered by exponential function.
The modification of this parameter has been demonstrated in the following:
1 Optimal Boosting Label Weighting Extreme Learning Machine … 9
r p (t + 1) = 4 ∗ r p (t) ∗ 1 − r p (t) , r p (t) ∈ (0, 1), p = 1, 2, (1.16)
−t
− tmax
w(t) = wmin + (wmax − wmin ) ∗ e −t
, (1.17)
whereas w min and w max represent the minimal and maximal of the inertia weights
and t max indicate the maximal amount of iterations.
The OBWELM-MDC technique has been simulated using three datasets, namely
depression, anxiety, and stress. All three dataset has 938 instances, 7 attributes, and
5 classes. Figure 1.2 illustrates the different DAS scales.
Figure 1.3 exhibits the confusion matrix of the OBWELM-MDC technique on
the test depression dataset. The figure reported that the OBWELM-MDC technique
has identified 707 instances under Normal, 59 instances into Mild, 59 instances into
Moderate, 19 instances into Severe, and 9 instances into Extreme classes.
Table 1.1 and Fig. 1.4 provide the classification result analysis of the OBWELM-
MDC technique on the depression dataset. The OBWELM-MDC technique has clas-
sified the samples into Normal class with accu y , precn , recal , spec y and Fscore of 93.28,
98.47, 93.15, 93.85, and 95.73%. Besides, the OBWELM-MDC system has classi-
fied the samples into Moderate class with accu y , precn , recal , spec y and Fscore of
97.44, 84.29, 81.94, 98.73, and 83.10%. Simultaneously, the OBWELM-MDC algo-
rithm has classified the samples into Extreme class with accu y , precn , recal , spec y
and Fscore of 99.25, 69.23, 75, 99.57, and 72%.
Figure 1.5 portrays the confusion matrix of the OBWELM-MDC method on the
test anxiety dataset. The figure stated that the OBWELM-MDC methodology has
identified 737 instances under Normal, 37 instances into Mild, 62 instances into
Moderate, 15 instances into Severe, and 3 instances into Extreme classes.
Table 1.2 and Fig. 1.6 offer the classification result analysis of the OBWELM-
MDC system on the anxiety dataset. The OBWELM-MDC technique has classified
the samples into Normal class with accu y , precn , recal , spec y and Fscore of 92.64,
98.79, 92.47, 93.62, and 95.53%.
Also, the OBWELM-MDC approach has classified the samples into Moderate
class with accu y , precn , recal , spec y and Fscore of 97.76, 82.67, 88.57, 98.50, and
85.52%. Concurrently, the OBWELM-MDC algorithm has classified the samples
into Extreme class with accu y , precn , recal , spec y and Fscore of 99.79, 75, 75, 99.89,
and 75%.
Figure 1.7 defines the confusion matrix of the OBWELM-MDC approach on
the test stress dataset. The figure obvious that the OBWELM-MDC technique has
identified 328 instances under Normal, 405 instances into Mild, 46 instances into
Moderate, 58 instances into Severe, and 25 instances into Extreme classes.
Table 1.3 and Fig. 1.8 give the classification result analysis of the OBWELM-
MDC method on the stress dataset. The OBWELM-MDC algorithm has classified the
samples into Normal class with accu y , precn , recal , spec y and Fscore of 95.32, 96.47,
90.86, 98, and 93.58%. Followed by, the OBWELM-MDC system has classified the
samples into Moderate class with accu y , precn , recal , spec y , and Fscore of 95.74, 63.89,
75.41, 97.11, and 69.17%. At last, the OBWELM-MDC approach has classified the
samples into Extreme class with accu y , precn , recal , spec y , and Fscore of 98.96, 71.43,
100, 98.93, and 83.33%.
Finally, a comparative result analysis of the OBWELM-MDC technique with other
recent methods is given in Table 1.4 [19, 20]. From the table, it can be observed that
the SVM and KNN models have obtained minimal classification performance. At the
same time, the DT, RF, and NB models have demonstrated moderate classification
results. However, the OBWELM-MDC technique has resulted in superior outcomes
over the other techniques. Therefore, it can be utilized as a proficient tool for the
detection and classification of mental disorders.
1.4 Conclusion
This article has designed an effective OBWELM-MDC technique for mental disorder
prediction and classification. The OBWELM-MDC technique follows a three-
stage process, namely pre-processing, BWELM-based prediction, and CSPSO-based
parameter tuning. In addition, the parameter tuning of the BWELM model takes place
by the use of CSPSO, where the inertia weight and acceleration coefficient of the
PSO algorithm are modified via logistic chaotic map. The utilization of CSPSO algo-
rithm has improved the predictive performance of the BWELM model. The exper-
imental result analysis of the OBWELM-MDC technique takes place using bench-
mark dataset, and the results are examined under several measures. The experimental
results showcased that OBWELM-MDC technique has accomplished maximum
predictive outcomes over the other methods. Therefore, the OBWELM-MDC tech-
nique can be utilized as an effective tool to predict and classify mental disorders. As
a part of future scope, the OBWELM-MDC technique can be extended to the design
of feature selection and feature reduction methodologies.
14 E. Laxmi Lydia et al.
Table 1.4 Comparative analysis of OBWELM-MDC technique in terms of different measures with
existing approaches
Classifier Dataset Accuracy Error Precision Recall Specificity F1
Score
Decision tree Anxiety 73.30 26.70 45.80 53.20 92.30 49.20
Depression 77.80 22.20 73.10 71.40 90.90 72.30
Stress 62.80 37.20 59.90 58.50 90.00 59.20
Random forest Anxiety 71.40 28.60 43.10 51.00 91.90 47.00
Depression 79.80 20.20 77.10 67.80 91.00 76.60
Stress 72.30 27.70 73.10 69.20 92.80 71.10
Naïve Bayes Anxiety 73.30 26.70 45.90 54.20 92.40 49.70
Depression 85.50 14.50 75.20 81.00 91.70 74.60
Stress 74.20 25.80 54.80 56.80 93.40 55.80
SVM model Anxiety 67.80 32.20 40.30 50.40 91.40 44.80
Depression 80.30 19.70 76.00 71.60 90.80 76.50
Stress 66.70 33.30 67.20 63.10 92.10 65.10
K nearest Anxiety 69.80 30.20 44.90 53.00 91.30 52.70
neighbour Depression 72.10 27.90 75.00 63.40 89.20 68.70
Stress 71.40 28.60 71.90 68.20 92.10 70.00
OBWELM-MDC Anxiety 96.42 3.58 70.45 82.98 97.04 75.10
Depression 96.37 3.63 72.18 82.47 97.09 76.19
Stress 95.84 4.16 77.77 87.84 97.42 81.88
References
1. Librenza-Garcia, D., Kotzian, B.J., Yang, J., Mwangi, B., Cao, B., Lima, L.N.P., Bermudez,
M.B., Boeira, M.V., Kapczinski, F., Passos, I.C.: The impact of machine learning techniques
in the study of bipolar disorder: a systematic review. Neurosci. Biobehav. Rev. 80, 538–554
(2017)
2. Chancellor, S., Baumer, E.P., De Choudhury, M.: Who is the “human” in human-centered
machine learning: the case of predicting mental health from social media. In: Proceedings of
the ACM on Human-Computer Interaction, vol. 3(CSCW), pp.1–32 (2019)
3. Mohr, D.C., Zhang, M., Schueller, S.M.: Personal sensing: understanding mental health using
ubiquitous sensors and machine learning. Annu. Rev. Clin. Psychol. 13, 23–47 (2017)
4. Thabtah, F.: Machine learning in autistic spectrum disorder behavioral research: a review and
ways forward. Inform. Health Soc. Care 44(3), 278–297 (2019)
5. Dwyer, D.B., Falkai, P., Koutsouleris, N.: Machine learning approaches for clinical psychology
and psychiatry. Annu. Rev. Clin. Psychol. 14, 91–118 (2018)
6. Bzdok, D., Meyer-Lindenberg, A.: Machine learning for precision psychiatry: opportunities
and challenges. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 3(3), 223–230 (2018)
7. Su, C., Aseltine, R., Doshi, R., Chen, K., Rogers, S.C., Wang, F.: Machine learning for suicide
risk prediction in children and adolescents with electronic health records. Transl. Psychiatry
10(1), 1–10 (2020)
1 Optimal Boosting Label Weighting Extreme Learning Machine … 15
8. Payrovnaziri, S.N., Chen, Z., Rengifo-Moreno, P., Miller, T., Bian, J., Chen, J.H., Liu, X., He,
Z.: Explainable artificial intelligence models using real-world electronic health record data: a
systematic scoping review. J. Am. Med. Inform. Assoc. 27(7), 1173–1185 (2020)
9. Jesmin, S., Kaiser, M.S., Mahmud, M.: Towards artificial intelligence driven stress monitoring
for mental wellbeing tracking during COVID-19. In: 2020 IEEE/WIC/ACM International Joint
Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), pp. 845–851.
IEEE (2020, December)
10. Nahian, M., Ghosh, T., Uddin, M.N., Islam, M., Mahmud, M., Kaiser, M.S.: Towards artifi-
cial intelligence driven emotion aware fall monitoring framework suitable for elderly people
with neurological disorder. In: International Conference on Brain Informatics, pp. 275–286.
Springer, Cham (2020, September)
11. Deshpande, M., Rao, V.: Depression detection using emotion artificial intelligence. In: 2017
International Conference on Intelligent Sustainable Systems (ICISS), pp. 858–862. IEEE (2017,
December)
12. Ćosić, K., Popović, S., Šarlija, M., Kesedžić, I., Jovanovic, T.: Artificial intelligence in predic-
tion of mental health disorders induced by the COVID-19 pandemic among health care workers.
Croat. Med. J. 61(3), 279 (2020)
13. Li, X., Zhou, P., Wu, J., Shanthini, A., Vadivel, T.: Research on artificial intelligence learning
system based on psychological knowledge to adjust anxiety and depression. Behav. Inform.
Technol. 1–13 (2020)
14. Pabreja, K., Singh, A., Singh, R., Agnihotri, R., Kaushik, S., Malhotra, T.: Stress predic-
tion model using machine learning. In: Proceedings of International Conference on Artificial
Intelligence and Applications, pp. 57–68. Springer, Singapore (2021)
15. Yu, H., Sun, C., Yang, X., Yang, W., Shen, J., Qi, Y.: ODOC-ELM: optimal decision outputs
compensation-based extreme learning machine for classifying imbalanced data. Knowl.-Based
Syst. 92, 55–70 (2016)
16. Cheng, K., Gao, S., Dong, W., Yang, X., Wang, Q., Yu, H.: Boosting label weighted extreme
learning machine for classifying multi-label imbalanced data. Neurocomputing 403, 360–370
(2020)
17. Yu, H., Sun, C., Yang, X., Zheng, S., Wang, Q., Xi, X.: LW-ELM: a fast and flexible cost-
sensitive learning framework for classifying imbalanced data. IEEE Access 6, 28488–28500
(2018)
18. Zhang, Z., Hong, W.C.: Application of variational mode decomposition and chaotic grey wolf
optimizer with support vector regression for forecasting electric loads. Knowl.-Based Syst.
228, 107297 (2021)
19. Mary, S.A., Jabasheela, L.: Ant colony optimization based feature selection and data
classification for depression anxiety and stress. IIOAB J. 9(2), 121–128 (2018)
20. Garg, S., Tigga, N.P.: Predicting anxiety, depression and stress in modern life using machine
learning algorithms. Procedia Comput. Sci. 167, 1258–1267 (2020)