An approach of cervical cancer diagnosis using class weighting and oversampling with Keras

TELKOMNIKA Telecommunication Computing Electronics and Control
Vol. 21, No. 1, February 2023, pp. 142~149
ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v21i1.24240  142
Journal homepage: http://telkomnika.uad.ac.id
An approach of cervical cancer diagnosis using class weighting
and oversampling with Keras
Hieu Le Ngoc, Khanh Vo Pham Huyen
Department of Computer Science, Faculty of Information Technology, Ho Chi Minh City Open University,
Ho Chi Minh City, Vietnam
Article Info ABSTRACT
Article history:
Received Jun 24, 2021
Revised Nov 06, 2022
Accepted Nov 16, 2022
Globally, cervical cancer caused 604,127 new cases and 341,831 deaths in
2020, according to the global cancer observatory. In addition, the number of
cervical cancer patients who have no symptoms has grown recently.
Therefore, giving patients early notice of the possibility of cervical cancer is
a useful task since it would enable them to have a clear understanding of
their health state. The use of artificial intelligence (AI), particularly in
machine learning, in this work is continually uncovering cervical cancer.
With the help of a logit model and a new deep learning technique, we hope
to identify cervical cancer using patient-provided data. For better outcomes,
we employ Keras deep learning and its technique, which includes class
weighting and oversampling. In comparison to the actual diagnostic result,
the experimental result with model accuracy is 94.18%, and it also
demonstrates a successful logit model cervical cancer prediction.
Keywords:
Cervical cancer diagnosis
Class weighting
Logit model
Machine learning
Oversampling
Prediction of cervical cancer This is an open access article under the CC BY-SA license.
Corresponding Author:
Hieu Le Ngoc
Department of Computer Science, Faculty of Information Technology, Ho Chi Minh City Open University
Room 605, 35-37 Ho Hao Hon Street, District 1, Ho Chi Minh City, Vietnam
Email: hieu.ln@ou.edu.vn
1. INTRODUCTION
With roughly 10 million deaths each year, cancer is one of the main causes of death in the globe.
According to globocan statistics [1], the world has about 1/6 of deaths due to cancer in 2020. The disease
manifests itself at a very late stage and the patients who are lacking diagnosis and treatment are very common.
The cervix, a woman’s opening from the vagina to the uterus, is where cervical cancer grows. It is the 4th
disease of cancer based on the number of new cases and the 6th
disease of cancer in the number of deaths among
women worldwide [2]. Women are most likely to get cervical cancer developing in their 30 s or older, it is the
average age of 48−52 years old. Although the disease causes great damage to the uterus, this disease progresses
silently for a long time (from 5 to 20 years) and its symptoms are quite faint, easy to get confused with other
gynecological diseases. So, it is difficult to detect this kind of disease in the early stages. In the late stages,
the patient must undergo a partial or complete removal of the uterus, which directly affects the reproductive
organs and motherhood of the woman. At the same time, artificial intelligence (AI) [3] is being applied to
diagnose diseases and manage all kinds of health problems. Researching on applying AI, especially machine
learning and deep learning techniques in cancer diagnosis, are very popular and developing in recent times.
From these issues above, we realize that building a model to diagnose cervical cancer is a practical
work. It can assist people in learning about their risk factors or potential for developing cancer from an early
stage. Therefore, they can take appropriate actions and control in time.
Studies on cervical cancer diagnosis through machine learning, most of which focus on the dataset on the
causes of cervical cancer risk, were observed at the Universitario de Caracas Hospital in Caracas, Venezuela [4].

TELKOMNIKA Telecommun Comput El Control 
An approach of cervical cancer diagnosis using class weighting and … (Hieu Le Ngoc)
143
When studying this dataset, we found that this data set is imbalanced and has a large bias, so it needs to apply
many methods and techniques to process. While Keras [5] is a library written in python language that
possesses many good supports for building deep learning models, we considered combining Keras deep
learning libraries in handling imbalanced datasets [6], [7]. Based on deep learning we propose using a logit
model built from the logistic regression algorithm a binary classification algorithm to diagnose a person with
cervical cancer or not, then apply class weighting and oversampling techniques and evaluate its accuracy and
performance through evaluation metrics like accuracy, precision, recall, F1 score. With the aid of these
methods, it can forecast cervical cancer by learning from the unbalanced data. It is feasible that no studies
have yet applied using this kind of method. Therefore, we would like to propose this approach in the detail in
the next section.
The recent studies’ machine learning algorithms are the premise and foundation for the inspiration
from this paper’s research direction, such as the research of Ijaz et al. [8] in 2020 with a model for predicting
cervical cancer using risk factors [4] as inputs. Utilizing noise and isolation forests with density-based spatial
clustering of applications, they eliminated outliers, using the synthetic minority over-sampling technique
(SMOTE) to balance the data, used the random forest (RF) as the classifier for the model. The model is built
with 4 cases that combined the above methods, and finally RF performed the best among all testing models. With
the same aims, the next study from Alsmariy et al. [9] in 2020 also used [4] to build a categorical model through
SMOTE to balance the dataset, principal component analysis to increase model performance and stratified
technique k-fold cross-validation to prevent overfitting problems combined with a trio of classifiers: decision tree
(DT), logistic regression, and RF. The results show that the high reliability of the model. Alam et al. [10] also used
SMOTE and a stratified 10-fold cross-validation technique to predict cervical cancer in 2019. The study using
three classifications: boosted DT, decision forest, and decision jungle algorithm on [4] which is balanced.
The above classifiers with the advantages of requiring fewer data to train as well as fast construction and
classification with high accuracy. The paper of Abisoye et al. [11] in 2019 proposed a study to predict
cervical cancer by combining a genetic algorithm (GA) to select the fittest features for the classification
process and a support vector machine (SVM) to build the classifier. The dataset [4] was processed to
eliminate noise using the min-max standardized technique, then the suitable features were selected through
the GA method. Finally, SVM was used in training the system using prepared data for cervical cancer
classification.
Two articles by Nithya and Ilango [12] and Mehmood et al. [13] in 2019 build some classifier models
based on machine learning algorithms. They use the feature selection technique to select features and the k-fold
cross-validation method to prevent overfitting. In the first one article the accuracy of the model increased from
97% to 100%, and in the second get higher metrics of area under the curve (AUC) and F1-score. This paper
shows that the better feature, the better model.
Besides cervical cancer, we also consider other cancer that has a similar concept to the article on
cervical cancer about how to use the risk factors dataset as input and methods and techniques to build
classification models. For example, the paper of Maleki et al. [14] in 2020, has applied k-nearest neighbors
(kNN) combined with a genetic algorithm to select risk features for lung cancer classification. The model is
built with 4 methods: DT, kNN without GA (𝑘 = 6, 10), and kNN with GA (𝑘 = 6). The result with the accuracy
of the classification increased to 100% after applying the above method. Or research by Khalil et al. [15]
in 2020, is about the development of a fuzzy expert system that is a method for predicting lung cancer with
the data is its symptoms. The generalization of cancer prediction paper by Maliha et al. [16] in 2019 used the
Naïve Bayes, k-NN, and J48 algorithm combined with 10-fold cross-validation to predict 9 types of popular
cancers with the dataset containing 61 attributes about symptoms. The Weka tool can then be used to assess
the precision of cancer data sets. Finally, a review paper [17] of 30 research papers in 2019, the authors
surveyed some past research papers and compared the accuracy of different machine learning (ML)
algorithms for cancer prediction depending on given datasets or their properties. Some articles used some
classifications such as SVM, RF, Naïve Bayes, DT, KNN, fuzzy neural network.
This paper is organized in 5 sections. The fisrt section talks about the cervical cancer situation,
the machine learning techniques implemented in this article, and the most related works on the topic. Section 2
is our proposed work which uses the logit model with the Keras technique to handle imbalanced data, introduce
the dataset, and the result of the pre-processing data. The experiment’s findings would be presented in section 4.
Section 5 concludes by summarizing the study’s overall findings.
2. PROPOSED METHOD
2.1. Theoretical basis
Keras is a library written in python, which provides the application programming interface (API) for
deep learning. It is built upon the well-known machine learning framework, TensorFlow. Keras helps us in
building logit models easily and dealing with imbalanced datasets [18] by using deep learning on its own.

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 1, February 2023: 142-149
144
We have numerous options for dealing with imbalanced data in TensorFlow and Keras [7]. But in this paper,
we are going to use two main techniques to handle imbalanced data are oversampling and class weights.
Also, receiver operating characteristic (ROC) techniques [19] which are used to avoid overfitting in the
model are dropout regularization and early stopping.
The metrics are used in this article from a confusion matrix for a binary classifier. The matrix is a table
where each column denotes a predicted class and each row denotes an actual class, as shown in Table 1.
In addition, we also use accuracy, precision, recall, ROC curve, and precision-recall curve (PRC) [20], [21].
Table 1. The confusion matrix
Predicted class actual class Positive Negative
Positive True positive (TP) False negative (FN)
Negative False positive (FP) True negative (TN)
2.2. Model building process
1) Input: patient data
The input of the proposal is the data of a patient. The input data will be handled via the predictive model
of the classifier. This model will use the dataset [4] as a training dataset to predict cancer. The essential
information of the data is the risk factor features for cervical cancer. These attributes will be detailed
described in the next section of this article.
2) Predictive model (building classifier)
The process of building a predictive model of cervical cancer in this proposed method includes 3 steps.
The first is to process the data, the second is to build the model, and the last is to predict and evaluate
the results. The specific work of each step is described in Figure 1.
− Step 1: firstly, from the data set of risk factors for cervical cancer [4] we process the data to get a
better dataset. In this step, we should clean the data to have better value and it can be fit with the
computation. We also need to create training and test sets from the dataset. Finally, the data is
normalized for better handling. The step is in detail as: 1) to clean data: the dataset has many
missing values, in this case, we used the dataset’s statistics as foundation, removed some attributes
that were not important or having too many missing values; 2) to split data: we split the dataset
into train, test, validation sets; and 3) to normalize data: after splitting the dataset, the training set
will be normalized by using the standardscaler [22] technique. By using this method, the mean and
standard deviation will both be set to 0 and 1.
− Step 2: after completing the data processing, we have a complete and normalized dataset. With this
new dataset, we will proceed to build a predictive model for the problem of binary classification.
We would apply the following two methods to increase the reliability: 1) to apply class weights
technique: this will utilize a parameter to pass the Keras weights for each class, this is to make the
model “become more attentive” to components with a low positive rate; and 2) to apply
oversampling technique: This will expand the training set’s sample size by expanding the minority
class’s sample size.
− Step 3: we predict the results with the test set and evaluate the accuracy of the model.
3) Output: conclude the patient is having cancer or not, with its percentage accuracy.
Figure 1. Model building process
2.3. Risk factors cervical cancer dataset
The risk factor cervical cancer dataset used in this article is from the University of California, Irvine
(UCI) machine learning repository [23]. The dataset was observed at Hospital Universitario de Caracas in
Caracas, Venezuela in 2017. The dataset includes 858 patients’ demographic data, lifestyle choices, and
previous medical records. Many omitted values exist in this dataset because some patients have decided not
to respond a few questions due to their privacy concerns. Table 2 contains the information describing the data
type, and how many values are missing for each attribute.

145
The research’s data collection of cervical cancer risk factors is a small dataset and imbalanced.
The attribute diagnosed (Dx): cancer is used to classify cancer with the number of patients diagnosed with
cancer being very small, a statistic is presented in Figure 2. With a sample of 858 patients, only 18 people
were diagnosed with cancer, respectively 2.1% of the total number of patients. This highly imbalanced effect
will affect the model’s predicted quality. In addition, Table 1 also shows that the quantity of missing values
in the dataset is very high (up to 26 out of 36 missing value properties) with the largest 2 missing attributes
being 787 and 18 missing attributes over 100 values.
Table 2. Attributes of the cervical cancer dataset
No. Attribute Type No. of
missing
value
No. Attribute Type No. of
missing
value
1 Age Integer 0 19 STDs: pelvic inflammatory disease Bool 105
2 Number of sexual partners Integer 26 20 STDs: genital herpes Bool 105
3 First sexual intercourse Integer 7 21 STDs: molluscum contagiosum Bool 105
4 Num of pregnancies Integer 56 22 STDs: acquired immunodeficiency
syndrome (AIDS)
Bool 105
5 Smokes Bool 13 23 STDs: human immunodeficiency
virus (HIV)
Bool 105
6 Smokes (years) Float 13 24 STDs: hepatitis B Bool 105
7 Smokes (packs/year) Integer 13 25 STDs: human papillomavirus (HPV) Bool 105
8 Hormonal contraceptives Bool 108 26 STDs: number of diagnoses Integer 0
9 Hormonal contraceptives (years) Float 108 27 STDs: time since first diagnosis Integer 787
10 IUD (intrauterine device) Bool 117 28 STDs: time since last diagnosis Integer 787
11 IUD (years of intrauterine device) Float 117 29 Dx: cancer Bool 0
12 Sexually transmitted diseases
(STDs)
Bool 105 30 Dx: cervical intra-epithelial (CIN) Bool 0
13 STDs (number) Integer 105 31 Dx: HPV Bool 0
14 STDs: condylomatosis Bool 105 32 Dx Bool 0
15 STDs: cervical condylomatosis Bool 105 33 Hinselmann Bool 0
16 STDs: vaginal condylomatosis Bool 105 34 Schiller Bool 0
17 STDs: vulvo-perineal
condylomatosis
Bool 105 35 Cytology Bool 0
18 STDs: syphilis Bool 105 36 Biopsy Bool 0
Figure 2. The small fraction of positive samples (cancer diagnosed)
2.4. Data processing
1) Clean data
− Remove the 6 diagnostic attributes: Dx: CIN, Dx: HPV, Dx: hinselmann, schiller, cytology,
biopsy. However, we keep the Dx: cancer attribute as the label for classify process.
− Remove the 2 attributes with many missing values: STDs: time since first diagnosis and STDs:
time since last diagnosis.
− Missing value: for attributes of true/false meaning, we replace missing values with a value of “0”.
For attributes that have a quantity meaning (numeric type), We substitute missing values with the
characteristics’ average values, where appropriate. After cleaning, the new dataset has 27
attributes, 26 attributes are risk factors, and 1 attribute is for classifying. All the attributes do not
contain any missing value.

 ISSN: 1693-6930
146
2) Split data
Three subsets of the dataset were created: training set, validation set, and testing set. Specifically, with the
sample of 858 tuples, the training set and testing set are separated from the dataset. The testing set is 20%
of the dataset. The training set makes up the remainder of the dataset, which takes 80%. This is a large
training set, and it will be divided into 2 more subsets, the validation set and the smaller training set.
The validation set also is 20% of this large training set. Each subset is divided into labels set (containing
the model’s classification class, the Dx: cancer attribute) and the features set (containing 26 attributes
about cervical cancer risk factors). As a result, we get the shape of the labels and features of the training,
validation, and test set respectively as (548), (138), (172), (548, 26), (138, 26), (172, 26).
3) Normalize data
We use StandardScaler to normalize the input features. The standard deviation will be set to 1 and the
mean to 0. This technique is provided by Keras and it is simply to use.
3. RESULTS AND DISCUSSION
3.1. Baseline model
Create a baseline model [24] by sequential [25] with 3 layers: dense, dropout (hidden layer), dense 1
(as output). The hidden layer relates to an input layer, a dropout layer is used to lessen overfitting, and an
output sigmoid layer is the ratio that the patient gets cancer. The model uses epochs = 40, batch size = 50 and
total params is 449. In addition, we also set the bias to reduce the loss of the model. We used Keras to
evaluate the baseline model with the testing set, with a sample of 172 patients, the model correctly predicted
170 non-cancer patients (true negatives) and failed to predict 2 cancer patients (false negatives), 98.8% in
accuracy. In this case, the results show that there are relatively few false predictions, which means that there
are relatively few cases of the disease that go undetected. However, we may want to have fewer false
negatives despite the number of false positives is rising. This compromise may be more appropriate since it is
safer to have fewer false negatives than to have more false positives. The results are shown by the ROC and
area under precision recall curve (AUPRC) charts [26].
The Figure 3(a) shows a low false-positive rate of around 2.5%. It is known that the testing set has
errors, but the training set does not. In Figure 3(b), the precision and recall distributions of the baseline model
are shown. For the training set, if the precision decreases, the recall will increase, the curve of testing
baseline set is a convex and concave polygon line due to the imbalanced dataset. This shows that the dataset
is not good, and we need to use some other techniques to ameliorate the fitness of the model.
(a) (b)
Figure 3. The baseline model results with (a) ROC chart and (b) AUPRC chart
3.2. Model improvement with class weighting
To improve the fit of the model, we use the class weighting technique provided by Keras. After
setup and calculation according to the guide of Keras, the weights of the model with classes 0 and 1 are 0.51
and 23.83, respectively. With the sample of 172 tuples, the model has predicted 160 non-cancer patients (true
negatives), 10 cancer patients but incorrectly (false positives) and 2 patient cancers (true positives) with an
accuracy is 94.18%. Although this model incorrectly predicted up to 10 people who did not get cancer,
in exchange, it can find people with the disease. In this case, we determine whether the patient has this
disease is much more important than misdiagnosing.

147
After using class weighting, the model result is quite smooth and fits better with the dataset. Take a look
at Figure 4, Figure 4(a) shows that the ratio of false positive after using the class weighting of the testing set
has increased and was greater than 2.5% and approximately 3.0%. The Figure 4(b), it is so obvious that the
precision and recall distributions are more reasonable, the testing set after weighting is closer to the training set,
showing an increased model fitness.
(a) (b)
Figure 4. Comparing the baseline model result and weighted model result in: (a) ROC chart and (b) AUPRC chart
3.3. Model improvement with oversampling
In the training set, the quantity of positive tuples is 15 and the number of negative samples is 535.
The oversampling method has increased the number of positive samples of the training set so that it is equal
to the number of negative samples. Then, we will combine this new positive sample with the negative
sample, so we have a new training set that is increased in size to 1070 records. This oversampling model
predicted 134 non-cancer patients, 36 cancer patients but incorrectly, and correctly predicted 2 cancer
patients with an accuracy of 79.07%. The results of the model are not very satisfactory compared to the class
weighting method.
After applying the oversampling technique, the resampled model also seems smoother and more fit
with the data, compared to the baseline model. In the left chart of Figure 5, the false positive ratio of the
resampled model is the smallest among the 3 models. With the AUC point is the greatest among them, the
resampled model shows us the most trusted model. But in the right chart of Figure 5, the distribution curve of
precision and recall of the 3 models, shows us a clearer picture of these techniques. The lines are for the
resampled model, we can see these 2 lines seem to be parallel and this represents the fittest model. But the
distance between these 2 lines is the longest one among the 3 of them. With the application of the 2 techniques,
we can see they are better than the original one, but each of these 2 techniques also has its own pros and cons.
So far, in this experiment result, we can see what we should do when we use the class weighting or
oversampling technique.
(a) (b)
Figure 5. Comparing the baseline model result, weighted model result and resampled model result in:
(a) ROC chart and (b) AUPRC chart

 ISSN: 1693-6930
148
When applying the oversampling technique, the resampled model also seems smoother and more fit
with the data, compared to the baseline model. In Figure 5(a), the false positive ratio of the resampled model is
the smallest among the 3 models. With the AUC point is the greatest among them, the resampled model shows
us the most trusted model. But in Figure 5(b), the distribution of precision and recall curve of the 3 models,
shows us a clearer picture of these techniques. The lines are for the resampled model, we can see these 2 lines
seem to be parallel and this represents the fittest model. But the distance between these 2 lines is the longest
one among the 3 of them. With the application of the 2 techniques, we can see they are better than the
original one, but each of these 2 techniques also has its own pros and cons. In brief, in this experiment result,
we can see what we should do when we use the class weighting or oversampling technique.
4. CONCLUSION
This article is a research result about the logit model, imbalanced data, and some deep learning
techniques is provided by Keras. We proposed a cancer disease prediction system that is used the model built
by the imbalanced dataset which contains the risk factors of cervical cancer. With the 3 models with different
techniques (baseline, class weighting and oversampling), the program runs out the diagnostic result coincides
with the fact that 2 records correctly reached the highest rate of 94.18% with the weighted model. This initial
information can be a warning which will be the basis for the patient to have an objective view of their disease,
so they can make plans to care for and to take treatment that is appropriate. For further development, we aim to
collect the most suitable dataset or process the dataset fitter for better prediction. Beyond that, we also need to
study more about other deep learning techniques in imbalanced dataset, and some improved logit models.
ACKNOWLEDGEMENTS
We would like to express our big appreciation to Ho Chi Minh City Open University in Vietnam for
supporting us in our research on this topic.
REFERENCES
[1] “Assessing national capacity for the prevention and control of noncommunicable diseases: report of the 2019 global survey,”
World Health Organization, 2020. https://apps.who.int/iris/handle/10665/331452 (accessed Mar. 15, 2021).
[2] “Cancer fact sheets,” Cancer today; World Health Organization, 2021. https://gco.iarc.fr/today/fact-sheets-cancers (accessed Mar. 15,
2021).
[3] B. J. Copeland, “Artificial intelligence,” 2022. https://www.britannica.com/technology/artificial-intelligence (accessed Jul. 13,
2022).
[4] K. Fernandes, J. S. Cardoso, and J. Fernandes, “Transfer learning with partial observability applied to cervical cancer screening,”
2017, pp. 243–250, doi: 10.1007/978-3-319-58838-4_27.
[5] F. Chollet, “Keras.” 2015. https://github.com/fchollet/Keras (accessed Apr. 16, 2022).
[6] “Classification on Imbalanced data,” Tensorflow. https://www.tensorflow.org/tutorials/structured_data/imbalanced_data (accessed
Apr. 07, 2021).
[7] “10 techniques to deal with Imbalanced classes in machine learning.” Analytics Vidhya.
https://www.analyticsvidhya.com/blog/2020/07/10-techniques-to-deal-with-class-imbalance-in-machine-learning/ (accessed Nov. 03,
2022).
[8] M. F. Ijaz, M. Attique, and Y. Son, “Data-driven cervical cancer prediction model with outlier detection and over-sampling
methods,” Sensors, vol. 20, no. 10, 2020, doi: 10.3390/s20102809.
[9] R. Alsmariy, G. Healy, and H. Abdelhafez, “Predicting cervical cancer using machine learning methods,” International Journal of
Advanced Computer Science and Applications, vol. 11, no. 7, 2020, doi: 10.14569/IJACSA.2020.0110723.
[10] T. M. Alam, M. M. A. Khan, M. A. Iqbal, A. Wahab, and M. Mushtaq, “Cervical cancer prediction through different screening
methods using data mining,” International Journal of Advanced Computer Science and Applications, vol. 10, no. 2, 2019,
doi: 10.14569/IJACSA.2019.0100251.
[11] Abisoye O. A., Abisoye B. O., E. Ayobami, and K. Lawal, “Prediction of cervical cancer occurrence using genetic algorithm and
support vector machine,” 3rd International Engineering Conference (IEC 2019), 2019. [Online]. Available:
http://repository.futminna.edu.ng:8080/jspui/bitstream/123456789/11397/1/PredictionofCervicalCancer.pdf
[12] B. Nithya and V. Ilango, “Evaluation of machine learning based optimized feature selection approaches and classification
methods for cervical cancer prediction,” SN Applied Sciences, vol. 1, 2019, doi: 10.1007/s42452-019-0645-7.
[13] M. Mehmood, M. Rizwan, M. Gregus ml, and S. Abbas, “Machine learning assisted cervical cancer detection,” Frontiers in
Public Health, vol. 9, 2021, doi: 10.3389/fpubh.2021.788376.
[14] N. Maleki, Y. Zeinali, and S. T. A. Niaki, “A k-NN method for lung cancer prognosis with the use of a genetic algorithm for
feature selection,” Expert Systems with Applications, vol. 164, 2021, doi: 10.1016/j.eswa.2020.113981.
[15] A. M. Khalil, S. -G. Li, Y. Lin, H. -X. Li, and S. -G. Ma, “A new expert system in prediction of lung cancer disease based on
fuzzy soft sets,” Soft Computing, vol. 24, pp. 14179–14207, 2020, doi: 10.1007/s00500-020-04787-x.
[16] S. K. Maliha, R. R. Ema, S. K. Ghosh, H. Ahmed, M. R. J. Mollick, and T. Islam, “Cancer disease prediction using naive bayes,
K-Nearest neighbor and J48 algorithm,” in 2019 10th International Conference on Computing, Communication and Networking
Technologies (ICCCNT), 2019, pp. 1–7, doi: 10.1109/ICCCNT45670.2019.8944686.
[17] A. Kumar, R. Sushil, and A. K. Tiwari, “Machine learning based approaches for cancer prediction: A survey,” 2nd International
Conference on Advanced Computing and Software Engineering (ICACSE-2019), 2019, doi: 10.2139/ssrn.3350294.
[18] H. He and E. A. Garcia, “Learning from imbalanced data,” IEEE Transactions on Knowledge and Data Engineering, vol. 21,
no. 9, pp. 1263–1284, 2009, doi: 10.1109/TKDE.2008.239.

149
[19] J. Kukačka, V. Golkov, and D. Cremers, “Regularization for deep learning: A taxonomy,” arXiv, 2017. [Online]. Available:
https://arxiv.org/pdf/1710.10686.pdf
[20] Hossin M. and Sulaiman M.N, “A review on evaluation metrics for data classification evaluations,” International Journal of Data
Mining & Knowledge Management Process, vol. 5, no. 2, pp. 1–11, 2015, doi: 10.5121/ijdkp.2015.5201.
[21] J. Davis and M. Goadrich, “The relationship between Precision-Recall and ROC curves,” in Proceedings of the 23rd international
conference on Machine learning - ICML ’06, 2006, pp. 233–240, doi: 10.1145/1143844.1143874.
[22] “Sklearn.preprocessing.standardscaler,” scikit-learn, 2020. https://scikit-learn.org/stable/modules/generated/sklearn.
preprocessing.StandardScaler.html (accessed May 15, 2020).
[23] “UCI machine learning repository,” Irvine, CA: University of California School of Information and Computer Science, 2019.
[Online]. Available: https://archive.ics.uci.edu/ml/
[24] “Baseline model,” Science direct. https://www.sciencedirect.com/topics/computer-science/baseline-model (accessed Jun. 07,
2021).
[25] Fchollet, “The sequential model,” Keras. https://Keras.io/guides/sequential_model/ (accessed Apr. 20, 2020).
[26] W. J. Krzanowski and D. J. Hand, ROC curves for continuous data, New York: Taylor & Francis, 2009,
doi: 10.1201/9781439800225
BIOGRAPHIES OF AUTHORS
Hieu Le Ngoc has been working in IT industry as an IT System Architect since
2010. As now, he is working as an IT instructor for HCMC Open University (Vietnam). His
major study is about cloud computing, cloud efficiency for better service, AI’s application,
Machine Learning and Computer Linguistics. His minor studies are education, education in IT
line, Languages Teaching (Chinese & English), Business and Economic. He can be contacted at
email: hieu.ln@ou.edu.vn and ResearchGate: https://www.researchgate.net/profile/Hieu-Le-24.
Khanh Vo Pham Huyen is an IT student at Ho Chi Minh City Open University.
She is getting Bachelor of IT in September 2021. She has been starting to study about data
and machine learning from 2020. She is keen on medicine data and how to discover a disease
with data. She would go on with further studies and keep moving up to post-graduation.
She can be contacted at email: 1751010056khanh@ou.edu.vn and ResearchGate:
https://www.researchgate.net/profile/Khanh-Vo-Pham-Huyen.

An approach of cervical cancer diagnosis using class weighting and oversampling with Keras

More Related Content

Similar to An approach of cervical cancer diagnosis using class weighting and oversampling with Keras (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

An approach of cervical cancer diagnosis using class weighting and oversampling with Keras