Resampling and Hyperparameter Tuning for Optimizing Breast Cancer
Prediction Using Light Gradient Boosting
Kartika Handayani
1
, Erni
1
, Rangga Pebrianto
1
, Ari Abdilah
1
, Rifky Permana
1
and Eni Pudjiarti
2
1
Universitas Bina Sarana Informatika, Jakarta, Indonesia
2
Universitas Nusa Mandiri, Jakarta,Indonesia
Keywords:
Breast Cancer, Light Gradient Boosting, Resampling, Hyperparameter Tuning.
Abstract:
Breast cancer is the most frequently diagnosed cancer and the leading cause of death. The main cause of
breast cancer is mainly related to patients who inherit genetic mutations in genes. Early diagnosis of breast
cancer patients is very important to prevent the rapid development of breast cancer apart from the evolution of
preventive procedures. A machine learning (ML) approach can be used for early diagnosis of breast cancer.
In this study, testing was performed using the Wisconsin Diagnostic Breast Cancer Dataset, also known as
WDBC (Diagnostics) which consists of 569 instances with no missing values and has one target class attribute,
either benign (B) or malignant (M). Tests were carried out using the ROS, RUS, SMOTE, and SMOTE-Tomek
resampling techniques to see the effect of overcoming unbalanced data. Then tested with Light Gradient
Boosting and optimized to get the best results using hyperparameter tuning. The best results are obtained after
tuning the hyperparameter with accuracy 99.12%, recall 99.12%, precisions 99.13%, f1-score 99.13% and
AUC 0.988.
1 INTRODUCTION
Breast cancer is the most frequently diagnosed can-
cer and the leading cause of death from cancer with
a percentage of 11.6%. In 2018 there were around
2.1 million cases of breast cancer in the world. It is
the most frequently diagnosed cancer in most coun-
tries (154 out of 185) and is also the leading cause of
death from cancer in more than 100 countries (Bray
et al., 2018). The main cause of breast cancer is
mainly related to patients inheriting genetic mutations
in their genes (Majeed, 2014). Breast cancer has two
different classes, namely benign and malignant. Be-
nign tumors, usually known as non-cancerous, while
malignant tumors or known as cancer can damage
the surrounding tissue and if the patient is diagnosed
as malignant, the doctor will perform a biopsy to
determine the aggressiveness or severity of the tu-
mor(Khan et al., 2017).
Early diagnosis of breast cancer patients is very
important to prevent the rapid development of breast
cancer apart from the evolution of preventive proce-
dures(Sun, 2017) and aids in the speedy recovery and
reduces the chance of death (Badr et al., 2019). Diag-
nosis of breast cancer can be done manually by a doc-
tor, but it takes a longer time and must be very compli-
cated for doctors to apply this classification (Khuriwal
and Mishra, 2018). The incompleteness of relevant
data can also lead to human error in diagnosis (Za-
itseva et al., 2020). Although advances in the treat-
ment of breast cancer have led to a reduction in mor-
tality from breast cancer across all age groups, young
age remains a high-risk factor and has a low survival
rate (Lee and Han, 2014). Due to the severity of the
disease, a computer-assisted detection (CAD) system
using a machine learning (ML) approach is required
for the diagnosis of breast cancer(Omondiagbe et al.,
2019). Therefore improving the accuracy of identify-
ing breast cancer is a very important task (Badr et al.,
2020). Various methods can be applied to the classifi-
cation of breast cancer (Idris and Ismail, 2021) to dis-
tinguish between two types of breast tumors namely
Benign and Malignant (Khan et al., 2017).
Classification is a data mining process that aims to
divide data into classes to facilitate decision-making
because that is an important task in the medical field.
Many researchers have done research to predict breast
cancer.Research done by proposing Support Vector
Machine (SVM) on the WBCD dataset produces an
accuracy of 86.10% (Kumari, 2018). Research that
Handayani, K., Erni, E., Pebrianto, R., Abdilah, A., Permana, R. and Pudjiarti, E.
Resampling and Hyperparameter Tuning for Optimizing Breast Cancer Prediction Using Light Gradient Boosting.
DOI: 10.5220/0012446100003848
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Advanced Information Scientific Development (ICAISD 2023), pages 177-180
ISBN: 978-989-758-678-1
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
177

proposes Random Forest using the WBCD dataset ob-
tains 91.66% accuracy (Pyingkodi et al., 2020). Re-
search with K- Nearest Neighbors(KNN) using the
WBCD dataset obtained an accuracy of 92.57%.
Based on the research that has been done, this
study proposes a comparison of sampling methods
such as Random Under Sampling (RUS), Random
Over Sampling (ROS), and Synthetic Minority Over
Sampling Technique (SMOTE) (Ernawan et al., 2022)
to see the effect on the problem of unbalanced data.
Then do the optimization of the best model with
hyperpa- rameter tuning. The purpose of this study
is to get the best predictive results from the proposed
classification model. This research is expected to pro-
duce better accuracy, recall, precision, f-score, and
AUC than previous research and add to research con-
tributions related to the data used.
2 MATERIALS AND METHODS
2.1 Datasets
The Wisconsin Diagnostic Breast Cancer Dataset,
also known as WDBC (Diagnostics). The source of
this dataset is the University of Wisconsin. This data
set consists of 569 instances with no missing val-
ues and has one target class attribute, either benign
(B) or malignant (M). The predictive attributes of the
data set consisted of ten real-valued features calcu-
lated for each core, such as radius, texture, circumfer-
ence, area, smoothness, compactness, concavity, sym-
metry, and fractal dimension. Mean, standard error,
and radius (average of the three largest value read-
ings) were calculated for each lead leading data set
that has more than 32 attributes, including the non-
predictive attribute i.e. patient ID.
2.2 Preprocessing
Data preparation stages are carried out to ensure that
the dataset used for training and testing is quality data.
If the dataset used still has noise, the resulting model
will also be of low quality and will have bias.
2.2.1 Data Cleaning
Data pre-processing begins with data cleaning which
consists of deleting the ”id” column in the origi-
nal dataset (WBCD) and diagnostic dataset (WDBC).
Then change the feature class name (as the target
class) to a feature with the name ”diagnosis” for
the three datasets, this is done because if you don’t
change the feature class name it will be read as a func-
tion.
2.2.2 Label Encodig
Data pre-processing begins with data cleaning which
consists of deleting the ”id” column in the origi-
nal dataset (WBCD) and diagnostic dataset (WDBC).
Then change the feature class name (as the target
class) to a feature with the name ”diagnosis” for
the three datasets, this is done because if you don’t
change the feature class name it will be read as a func-
tion.
2.2.3 Split Data
Split data: 80% data training, 20% data testing, train-
ing data is used to build and train the model, and data
testing will be used for model evaluation and model
testing.
2.2.4 Resampling
The WDBC dataset with the target class ”benign” has
a total of 357 data and the target class ”malignant” has
a total of 212 data with a total of 569 instances. The
following is a visualization of the data distribution of
the WDBC dataset target class.
Figure 1: Class Distribution.
For the Random Under Sampling (RUS), most
of the category examples are discarded until an
even distribution of information is achieved (Dorn,
2021). The Random Over Sampling (ROS) algo-
rithm randomly replicates samples from the minor-
ity class(Rend
´
on et al., 2020). Oversampling can
be done by increasing the number of instances or
minority class samples based on the production of
new samples or repeated samples(Mohammed et al.,
2020). The Synthetic Minority Over Sampling Tech-
nique SMOTE generates an artificial sample of the
minority class by interpolating existing instances that
are very close to each other (Rend
´
on et al., 2020). For
the minority category in the information set, SMOTE
initially selects the minority class data instance at ran-
dom. Then, Ma’s k nearest neighbors are related to
ICAISD 2023 - International Conference on Advanced Information Scientific Development
178

the minority class, ID (Dorn, 2021). SMOTE with
Tomek Link balances knowledge and instant builds
of well-separated additional categories. during this
approach, any data instance that generates a link that
Tomek discards are of either the minority or the ma-
jority class (Dorn, 2021).
2.3 Model
2.3.1 Light Gradient Boosting
LightGBM uses gradient enhancement in its construc-
tion, but light GBM does not split the eigenvalues
individually, so it is necessary to calculate the split-
ting benefit of each eigenvalue. LightGBM algorithm
on the model to improve forecasting accuracy and ro-
bustness (Ju et al., 2019). Can find the optimal split
value (Su, 2020).
2.3.2 Hyperparameter Tuning
Hyperparameter tuning can be complete when the
data is large (Joy et al., 2016), showing more hyperpa-
rameters to tune, and models with carefully designed
structures imply that hyperparameters must be set too
tight ranges to reproduce the precision (Yu and Zhu,
2020).
2.3.3 GridSearch
To increase the optimal hyperparameter, the Grid-
SearchCV model taken from Scikit learn is used
[22]to get the best parameters. GridSearchCV im-
plements the fit and score methods. It also imple-
ments score samples, predict, predict proba, deci-
sion function, transform and inverse transform if im-
plemented in the estimator used (Pedregosa, 2011).
The parameter estimator used to implement this
method is optimized by cross-validation through grid
parameters.
3 RESULT AND DISCUSSION
The following tests are carried out using light gradient
boosting without using hyperparameter tuning.
Table 1: Testing Without Hyperparameter Tuning.
Resampling Accuracy Recall Precision F1-Score AUC
Without Re-
sampling
95.61% 90.48% 97.44% 93.83% 0.945
ROS 95.61% 90.48% 97.44% 93.83% 0.945
SMOTE 96.49% 92.86% 97.50% 95.12% 0.957
RUS 97.37% 95.24% 97.56% 96.36% 0.969
SMOTE
Tomek
94.73% 88.09% 97.36% 92.50% 0.934
The best results before testing without hyperpa-
rameter tuning were obtained by resampling RUS
with accuracy results 97.37%, recall 95.24%, preci-
sion 97.56%,f1-score 96.36%a nd AUC 0.934. Then
a search for hyperparameter tuning is performed with
a grid search based on the following table 2.
Table 2: Parameters Tested in the Hyperparameter tuning
Grid Search.
n estimators 100, 400, 10
min child weight 3, 20, 2
colsample bytree 0.4, 1.0
max deph 5, 15, 2
num leaves 8, 40
min child weight 10.30
learning rate 0.01,1
After searching for hyperparameter tuning, the
best parameters are obtained as follows
Table 3: Best Parameters.
’boosting type’ ’gbdt’,
’class weight’ none,
’colsample bytree’ 1.0,
’importance type’ ’split’,
’learning rate’ 0.1,
’max depth’ 15,
’min child samples’ 10,
’min child weight’ 20,
’min split gain’ 0.0,
’n estimators’ 400,
’n jobs’ -1,
’num leaves’ 31,
’objective’ none,
’random state’ none,
’reg alpha’ 0.0,
’reg lambda’ 0.0,
’silent’ true,
’subsamples’ 1.0,
’subsample for bin’ 200000,
’subsample freq’ 0
After the best parameter results were obtained, the
test was carried out again with the results in the fol-
lowing table
Table 4: Testing With Hyperparameter Tuning.
Resampling accuracy recall Precision F1-Score AUC
Without Re-
sampling
99.12% 99.12% 99.13% 99.13% 0.988
ROS 99.12% 99.12% 99.13% 99.13% 0.988
SMOTE 99.12% 99.12% 99.13% 99.13% 0.988
RUS 98.24% 98.24% 98.24% 98.24% 0.981
SMOTE
Tomek
96.49% 96.49% 96.68% 96.68% 0.952
The results using hyperparameter tuning have in-
Resampling and Hyperparameter Tuning for Optimizing Breast Cancer Prediction Using Light Gradient Boosting
179

creased in all tests. With the best results with the
same value on the test without resampling, ROS, and
SMOTE. The best results are obtained with the accu-
racy 99.12%, recall 99.12%, precisions 99.13%, f1-
score 99.13% and AUC 0.988.
4 CONCLUSIONS
This study conducted a test for breast cancer predic-
tion. The use of resampling techniques such as ROS,
SMOTE, RUS, and SMOTE-Tomek is done to over-
come unbalanced data. The use of hyperparameter
tuning using a grid search with light gradient boost-
ing results in an increase and optimization of results.
Obtain the best results with accuracy 99.12%, re-
call 99.12%, precisions 99.13%, f1-score 99.13% and
AUC 0.988. In further research, testing can be done
with other breast cancer datasets or with other meth-
ods.
REFERENCES
Badr, E., Abdulsalam, M., and Ahmed, H. (2020). The
impact of scaling on support vector machine in
breast cancer diagnosis. Int. J. Comput. Appl,
175(19):15–19,.
Badr, E., Salam, M., and Ahmed, H. (2019). Optimizing
support vector machine using gray wolf optimizer al-
gorithm for breast cancer detection.
Bray, F., Ferlay, J., Soerjomataram, I., Siegel, R., Torre, L.,
and Jemal, A. (2018). Global cancer statistics 2018:
Globocan estimates of incidence and mortality world-
wide for 36 cancers in 185 countries. CA. Cancer J.
Clin, 68(6):394–424,.
Dorn, M. (2021). Comparison of machine learning tech-
niques to handle imbalanced covid-19 cbc datasets.
PeerJ Comput. Sci, 7:670,.
Ernawan, F., Handayani, K., Fakhreldin, M., and Abbker, Y.
(2022). Light gradient boosting with hyper parameter
tuning optimization for covid-19 prediction. Int. J.
Adv. Comput. Sci. Appl, 13(8):514–523,.
Idris, N. and Ismail, M. (2021). Breast cancer disease clas-
sification using fuzzy-id3 algorithm with fuzzydbd
method: Automatic fuzzy database definition. PeerJ
Comput. Sci, 7:1–22,.
Joy, T., Rana, S., Gupta, S., and Venkatesh, S. (2016).
Hyperparameter tuning for big data using bayesian
optimisation. Proc. - Int. Conf. Pattern Recognit,
0:2574–2579,.
Ju, Y., Sun, G., Chen, Q., Zhang, M., Zhu, H., and
Rehman, M. (2019). A model combining convo-
lutional neural network and lightgbm algorithm for
ultra-short-term wind power forecasting. IEEE Ac-
cess, 7(c):28309–28318,.
Khan, R., Suleman, T., Farooq, M., Rafiq, M., and Tariq,
M. (2017). Data mining algorithms for classifica-
tion of diagnostic cancer using genetic optimization
algorithms. IJCSNS Int. J. Comput. Sci. Netw. Secur,
12(March):207–211,.
Khuriwal, N. and Mishra, N. (2018). Breast cancer diag-
nosis using deep learning algorithm. In Proc. - IEEE
2018 Int. Conf. Adv. Comput. Commun. Control Net-
working, ICACCCN 2018, pages 98–103,.
Kumari, M. (2018). The impact of news information on the
stock recommendation system: A survey.
Lee, H. and Han, W. (2014). Unique features of young age
breast cancer and its management. J. Breast Cancer,
17(4):301–307,.
Majeed, W. (2014). Breast cancer: Major risk factors and
recent developments in treatment. Asian Pacific J.
Cancer Prev, 15(8):3353–3358,.
Mohammed, R., Rawashdeh, J., and Abdullah, M. (2020).
Machine learning with oversampling and undersam-
pling techniques: Overview study and experimental
results. 11th Int. Conf. Inf. Commun. Syst. ICICS
2020, (May):243–248,.
Omondiagbe, D., Veeramani, S., and Sidhu, A. (2019). Ma-
chine learning classification techniques for breast can-
cer diagnosis. IOP Conf. Ser. Mater. Sci. Eng, 495(1).
Pedregosa, F. (2011). Scikit-learn: Machine learning in
python. J. Mach. Learn. Res, 12:2825–2830,.
Pyingkodi, M., M., M., Shanthi, S., Saravanan, T., Then-
mozhi, K., Nanthini, K., Hemalatha, D., and Dhivya
(2020). Performance study of classification algo-
rithms using the breast cancer dataset. Int. J. Futur.
Gener. Commun. Netw.
Rend
´
on, E., Alejo, R., Castorena, C., Isidro-Ortega, F., and
Granda-Guti
´
errez, E. (2020). Data sampling methods
to dealwith the big data multi-class imbalance prob-
lem. Appl. Sci, 10(4).
Su, Y. (2020). Prediction of air quality based on gradi-
ent boosting machine method. In Proc. - 2020 Int.
Conf. Big Data Informatiz. Educ. ICBDIE 2020, pages
395–397,.
Sun, Y. (2017). Risk factors and preventions of breast can-
cer. Int. J. Biol. Sci, 13(11):1387–1397,.
Yu, T. and Zhu, H. (2020). Hyper-parameter optimization:
A review of algorithms and applications.
Zaitseva, E., Levashenko, V., Rabcan, J., and Krsak, E.
(2020). Application of the structure function in the
evaluation of the human factor in healthcare. Symme-
try, 12(1):93,.
ICAISD 2023 - International Conference on Advanced Information Scientific Development
180