A Comparison of Undersampling, Oversampling, and SMOTE Methods for Dealing with Imbalanced Classification in Educational Data Mining

Wongvorachan, Tarid; He, Surina; Bulut, Okan

doi:10.3390/info14010054

Cited by 128 publications

(62 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…given by some literature is exactly 40% of the total sample size of positive cases [20,37,38], Of course, the SMOTE-NC also has limitations, although it has been proven to perform well in many fields. [39] Therefore, when we need to balance the categories of clinical problems, maybe RF plus SMOTE-NC is a better combination. Caution is advised in interpreting this conclusion, and further evidence is warranted for confirmation in the immediate future.…”

Section: Discussionmentioning

confidence: 99%

Development and Validation of a Machine learning Model for the Prediction of Short-Term Mortality among Patients with Sepsis: A Retrospective Cohort Study

Hu,

Ming,

Zhu

et al. 2024

Preprint

View full text Add to dashboard Cite

Background: Sepsis, a severe condition arising from the body's response to infection, leading to organ damage, is a major contributor to mortality in intensive care units. This study assesses and compares the predictive performance of six machine learning algorithms in forecasting in-patient mortality among sepsis patients. Results: There were 6,455 sepsis patients (5,165 for training and 1,290 for internal validation) in Medical Information Mart for Intensive Care IV v2.0 and 411 sepsis patients in eICU Collaborative Research Database v2.0 who met the inclusion criteria. The median age of the patients was 64 years and 61 years respectively. 42 features were retained for model building. The most important features for predicting mortality were urine output, Glasgow Coma Scale score, urine output, average body temperature, age, and lactate. The random forest algorithm performed the best on both the imbalanced and balanced datasets, achieving an AUC of 78.6% and 78.5%, respectively, on the external test set. Conclusions: This study showcases the applicability of machine learning for outcome prediction in sepsis patients, employing internal and multicenter external validation datasets. The application of the SHAP method enhances model interpretability by providing insights into global explanations, individual explanations, and the interactions between features.

show abstract

Section: Discussionmentioning

confidence: 99%

Development and Validation of a Machine learning Model for the Prediction of Short-Term Mortality among Patients with Sepsis: A Retrospective Cohort Study

Hu,

Ming,

Zhu

et al. 2024

Preprint

View full text Add to dashboard Cite

show abstract

“…SMOTE [24] is a classical oversampling method to solve the data imbalance problem. It uses the idea of linear interpolation to construct minority samples in the nearest neighborhood.…”

Section: Resampling Methodsmentioning

confidence: 99%

“…Thirdly, compared to other methods, evolutionary features achieve best results, but the corresponding feature dimensions are also relatively higher, especially for the M495 dataset. Fourthly, the predictive accuracy of New-All-2 is not only best but also has lower feature dimensions, outperforming the combined methods of Group1, Group2, and Group3: To deal with the problem of sample imbalance, four different sampling methods were employed for comparison-SMOTE [24], SMOTEENN [25], ADASYN [28], and SMOTETOMEK [29]. The results are presented in Table 7.…”

Section: Effect Of Various Feature Extraction Methodsmentioning

confidence: 99%

Ensemble Learning Method for Predicting Protein Submitochondrial Localization by Multi-type Feature Fusion

Ruan,

Liu,

Yang

et al. 2024

Preprint

View full text Add to dashboard Cite

Background The mitochondria exist in most eukaryotic cells and mainly function as energy producer. Different mitochondrial structures have different functional areas. Abnormal mitochondrial structures not only cause diseases but also promote aging. In this regard, accurate prediction of protein submitochondrial locations could enhance the understanding of disease mechanisms and promote drug design. However, current mainstream method mainly adopts single attribute features and barely considers unbalanced samples, so, we introduce a method to resolve these problems and improve prediction accuracy in this work. Results We combined diverse types of features that were extracted based on sequence information, evolutionary information, and physicochemical property to enrich feature expression. The integrated class prediction information of each feature description was based on a two-layer support vector machine ensemble model to eliminate the heterogeneity of the feature fusion. Furthermore, resampling method was utilized to settle the unbalanced class. The jackknife test was employed to validate and compare the overall prediction framework with existing methods on M317 and M983 datasets. Independent test was applied on M495 to prove the generalization of the model. Experimental results revealed that the prediction performance for multi-source feature integration was higher than that for homologous feature fusion, and it improved the accuracy of the small class sample and surpassed other methods in terms of overall prediction accuracy. Finally, local software for resampling was built for reader’s convenience. Conclusions The results show that our method can not only effectively handle the problem of imbalanced samples, but also has good predictive performance and stability. The datasets and codes can be freely obtained at https://github.com/ruanxiaoli/Ensemble-Learning.

show abstract

“…As expected, the brightest students are enrolled in the medicine program, so their GPA is high. However, the class imbalance problem in educational datasets could hamper the accuracy of predictive models as many of these models are designed on the assumption that the predicted class is balanced [ 27 ]. Therefore, to have an equitable class distribution for the target variable GPA, students were divided into two classes: “excellent,” having a GPA more than 4.25, and “good,” having a GPA less than 4.25 but more than 2.…”

Section: Methodsmentioning

confidence: 99%

“…The EDM community uses four major approaches: prediction models, structure discovery, relationship mining and discovery with models [ 26 ]. Educational data mining is capable of producing useful data-driven applications (e.g., early warning systems in schools or the prediction of students’ academic achievement) based on predictive models [ 27 ]. Researchers at King Khalid University (KKU) have used Orange data mining tool for EDM for detecting patterns and predicting academic performance of students using online courses offered through learning management systems [ 28 ].…”

Section: Introductionmentioning

confidence: 99%

Investigating Students’ Pre-University Admission Requirements and Their Correlation with Academic Performance for Medical Students: An Educational Data Mining Approach

2023

View full text Add to dashboard Cite

Medical education is one of the most sought-after disciplines for its prestigious and noble status. Institutions endeavor to identify admissions criteria to register bright students who can handle the complexity of medical training and become competent clinicians. This study aims to apply statistical and educational data mining approaches to study the relationship between pre-admission criteria and student performance in medical programs at a public university in Saudi Arabia. The present study is a retrospective cohort study conducted at the College of Computer Science, King Khalid University, Abha, Kingdom of Saudi Arabia between February and November 2022. The current pre-admission criterion is the admission score taken as the weighted average of high school percentage (HSP), general aptitude test (GAT) and standard achievement admission test (SAAT), with respective weights of 0.3, 0.3 and 0.4. Regression and optimization techniques have been applied to identify weightages that better fit the data. Five classification techniques—Decision Tree, Neural Network, Random Forest, Naïve Bayes and K-Nearest Neighbors—are employed to develop models to predict student performance. The regression and optimization analyses show that optimized weights of HSP, GAT and SAAT are 0.3, 0.2 and 0.5, respectively. The results depict that the performance of the models improves with admission scores based on optimized weightages. Further, the Neural Network and Naïve Bayes techniques outperform other techniques. Firstly, this study proposes to revise the weights of HSP, GAT and SAAT to 0.3, 0.2 and 0.5, respectively. Secondly, as the evaluation metrics of models remain less than 0.75, this study proposes to identify additional student features for calculating admission scores to select ideal candidates for medical programs.

show abstract

A Comparison of Undersampling, Oversampling, and SMOTE Methods for Dealing with Imbalanced Classification in Educational Data Mining

Cited by 128 publications

References 33 publications

Development and Validation of a Machine learning Model for the Prediction of Short-Term Mortality among Patients with Sepsis: A Retrospective Cohort Study

Development and Validation of a Machine learning Model for the Prediction of Short-Term Mortality among Patients with Sepsis: A Retrospective Cohort Study

Ensemble Learning Method for Predicting Protein Submitochondrial Localization by Multi-type Feature Fusion

Investigating Students’ Pre-University Admission Requirements and Their Correlation with Academic Performance for Medical Students: An Educational Data Mining Approach

Contact Info

Product

Resources

About