The DOI matches its Crossref record; 77 of the 83 references it cites were found in a registry or an index.
A systematic study of the class imbalance problem in convolutional neural networks
Open accessAbstract
In this study, we systematically investigate the impact of class imbalance on classification performance of convolutional neural networks (CNNs) and compare frequently used methods to address the issue. Class imbalance is a common problem that has been comprehensively studied in classical machine learning, yet very limited systematic research is available in the context of deep learning. In our study, we use three benchmark datasets of increasing complexity, MNIST, CIFAR-10 and ImageNet, to investigate the effects of imbalance on classification and perform an extensive comparison of several methods to address the issue: oversampling, undersampling, two-phase training, and thresholding that compensates for prior class probabilities. Our main evaluation metric is area under the receiver operating characteristic curve (ROC AUC) adjusted to multi-class tasks since overall accuracy metric is associated with notable difficulties in the context of imbalanced data. Based on results from our experiments we conclude that (i) the effect of class imbalance on classification performance is detrimental; (ii) the method of addressing class imbalance that emerged as dominant in almost all analyzed scenarios was oversampling; (iii) oversampling should be applied to the level that completely eliminates the imbalance, whereas the optimal undersampling ratio depends on the extent of imbalance; (iv) as opposed to some classical machine learning models, oversampling does not cause overfitting of CNNs; (v) thresholding should be applied to compensate for prior class probabilities when overall number of properly classified cases is of interest.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“A memory of the past is needed to rectify past classes' scores.”
From A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks (2020) · cites this paper as [12]
In the citing paper
“Here, we introduced a competitive method where classification bias in favor of new classes is reduced by using prior class probabilities [12].”
From A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks (2020) · cites this paper as [12]
In the citing paper
“Finally, the highly-discussed dimension target class balance pays tribute to the technical requirements of ML140,141,144,150,159.”
From The METRIC-framework for assessing data quality for trustworthy AI in medicine: A systematic review (2024) · cites this paper as [141]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 1×
“And for all the cases, the classes chosen to be minority were the ones with the lowest label value.”
BackgroundVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“A standard CNN architecture comprises fully connected layers and a number of blocks consisting of convolution kernels, activation function layer and max pooling [22, 23, 24].”
MethodImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
Scikit-learn: Machine Learning in Python
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Müller, A., Nothman, J., Louppe, G., et al. (2012) · arXiv · doi:10.48550/arxiv.1201.0490 · Cited by 71,016 · cited 1×
“We used a specific implementation of the ROC AUC available in scikit-learn python package [65].”
MethodGradient-based learning applied to document recognition
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998) · Proceedings of the IEEE · doi:10.1109/5.726791 · Cited by 58,805 · cited 2×
“Images The CNN model that we use for MNIST is the modern version of LeNet-5 [52].”
MethodImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 1×
“In our study, we used three benchmark datasets: MNIST [52], CIFAR-10 [53] and ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2012 [54].”
MethodSMOTE synthetic minority over-sampling technique
Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002) · Journal of Artificial Intelligence Research · doi:10.1613/jair.953 · Cited by 32,865 · cited 3×
“For classical machine learning models it was shown that oversampling can cause overfitting, especially for minority classes [33].”
BackgroundMethodDelving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
He, K., Zhang, X., Ren, S., & Sun, J. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.123 · Cited by 19,741 · cited 1×
“Network weights were initialized with Kaiming (also known as MSRA) initialization procedure [63].”
MethodVisualizing and understanding convolutional networks
Zeiler, M. D., & Fergus, R. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10590-1_53 · Cited by 15,839 · cited 1×
“This property allows CNNs to learn hierarchical representations [2].”
BackgroundBatch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate ShiftPreprint
Ioffe, S. (2024) · arXiv (Cornell University) · doi:10.57702/o9raffed · Cited by 15,273
Cited by
Sorted by citations
Items 801–900 of 1,001
Dong, R., Mou, L., Chen, M., Li, W., Tong, X., Yuan, S., Zhang, L., Zheng, J., Zhu, X. X., & Fu, H. (2023) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.01539 · Cited by 20
“To solve the class-imbalance problem, related works include data re-sampling [3, 4], loss re-weighting [22, 32, 18], margin modification [5, 35], and decoupled learning [20, 47].”
MethodCoquet, J., Bozkurt, S., Kan, K. M., Ferrari, M. K., Blayney, D. W., Brooks, J. D., & Hernandez‐Boussard, T. (2019) · Journal of Biomedical Informatics · doi:10.1016/j.jbi.2019.103184 · Cited by 19
Sparse Matrix Classification on Imbalanced Datasets Using Convolutional Neural Networks
Pichel, J. C., & Pateiro‐López, B. (2019) · IEEE Access · doi:10.1109/access.2019.2924060 · Cited by 19
“Methods for addressing imbalance for classical machine learning models have been studied for many years [5], [16], while recent works are mainly focused on CNNs [17].”
BackgroundFacial Action Units for Training Convolutional Neural Networks
Pham, T. T. D., & Won, C. S. (2019) · IEEE Access · doi:10.1109/access.2019.2921241 · Cited by 22
“Specific suggestions for resolving the imbalance problem have been made [9]: (i) For extreme ratio of imbalance and large portion of classes being minority, undersampling performs on a par with oversampling, (ii) Oversampling should be applied to the level that completely eliminates the imbalance, whereas the optimal…”
BackgroundInfluentialZhao, K., Liu, Y., Hao, S., Lu, S., Liu, H., & Zhou, L. (2021) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2021.3064316 · Cited by 19
“5 we can learn that there is a class imbalance problem [49] in BEAUTY.”
Classifying functional nuclear images with convolutional neural networks: a survey
Lin, Q., Man, Z., Cao, Y., Deng, T., Han, C., Cao, C., Zhang, L., Zeng, S., Gao, R., Wang, W., et al. (2020) · IET Image Processing · doi:10.1049/iet-ipr.2019.1690 · Cited by 20
Olivares, E., Ye, H., Herrero, A., Nia, B. A., Ren, Y., Donovan, R. P., & Flaviis, F. D. (2021) · Neurocomputing · doi:10.1016/j.neucom.2021.04.021 · Cited by 20
Artificial Intelligence-Enabled ECG Algorithm for the Prediction of Coronary Artery Calcification
Han, C., Kang, K., Kim, T. Y., Uhm, J., Park, J., Jung, I. H., Kim, M., Bae, S., Lim, H., & Yoon, D. (2022) · Frontiers in Cardiovascular Medicine · doi:10.3389/fcvm.2022.849223 · Cited by 20
“The dataset used in this study was a class imbalanced dataset and it has been established that class imbalance can have detrimental effects on classification performance, and the most widely used robust method for dealing with class imbalance is the oversampling of the minority class (24).”
Zhang, T. G., Zhu, C., Zhao, Y., Zhao, M., Wang, Z., Song, R., Meng, N., Sial, A. W., Diwan, A. D., Liu, J., et al. (2022) · Research Square · doi:10.21203/rs.3.rs-1655808/v1 · Cited by 19
Automatically weighted focal loss for imbalance learning
Mahmoodi, N., Shirazi, H. M., Fakhredanesh, M., & DadashtabarAhmadi, K. (2024) · Neural Computing and Applications · doi:10.1007/s00521-024-10323-x · Cited by 18
“A common data method of solving the imbalance issue is to resample, either by under-sampling such as (Random Under-Sampling) RUS, or by oversampling such as (Random Over-Sampling) ROS [3–5].”
Azqadan, E., Arami, A., & Jahed, H. (2025) · Journal of Magnesium and Alloys · doi:10.1016/j.jma.2025.07.019 · Cited by 18
Polat, H., Özerdem, M. S., Ekici, F., & Akpolat, V. (2021) · International Journal of Imaging Systems and Technology · doi:10.1002/ima.22558 · Cited by 18
“37 The overfitting problem may occur for the class, which constitutes the majority in the data distribution by the proposed classifier model.”
BackgroundChen, X., Liu, H., Xu, X., Zhang, L., Lin, T., Zuo, M., Huang, Y., Shen, R., Chen, D., & Deng, Y. (2021) · ACS ES&T Water · doi:10.1021/acsestwater.1c00037 · Cited by 18
Instance-Aware Plant Disease Detection by Utilizing Saliency Map and Self-Supervised Pre-Training
Kim, T., Kim, H., Baik, K., & Choi, Y. (2022) · Agriculture · doi:10.3390/agriculture12081084 · Cited by 21
“Even if a dataset considering all these cases is acquired, supervised-learning-based CNN models tend to be overfitted to a dominant class due to the class imbalance problem [20,21] between class categories, which deteriorates the ability of feature presentation and may result in redundant performance for classes with…”
BackgroundReal-time detection of electrical load anomalies through hyperdimensional computing
Wang, X., Flores, R. J., Brouwer, J., & Papaefthymiou, M. C. (2022) · Energy · doi:10.1016/j.energy.2022.125042 · Cited by 20
Dimensionality reduction techniques in structural and earthquake engineering
Hajibabaee, P., Pourkamali‐Anaraki, F., & Hariri‐Ardebili, M. A. (2023) · Engineering Structures · doi:10.1016/j.engstruct.2022.115485 · Cited by 19
Detection of Illegal Transactions of Cryptocurrency Based on Mutual Information
Zhao, K., Dong, G., & Bian, D. (2023) · Electronics · doi:10.3390/electronics12071542 · Cited by 19
Liu, S., Chen, H., Liu, P., Qin, F., & Fars, A. (2023) · International Journal of Hydrogen Energy · doi:10.1016/j.ijhydene.2023.04.091 · Cited by 18
RetractedGigapixel end-to-end training using streaming and attention
Dooper, S., Pinckaers, H., Aswolinskiy, W., Hebeda, K. M., Jarkman, S., van der Laak, J. A. W. M., & Litjens, G. J. S. (2023) · Medical Image Analysis · doi:10.1016/j.media.2023.102881 · Cited by 20
Zhu, Y., Li, Y., & Zhang, Q. (2023) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2023.3302472 · Cited by 18
Use Your Head: Improving Long-Tail Video Recognition
Perrett, T., Sinha, S., Burghardt, T., Mirmehdi, M., & Damen, D. (2023) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52729.2023.00239 · Cited by 22
“In order to encourage methods to train effectively on long-tail data, image-recognition benchmarks include multiple naturally-collected 1 [24] as well as curated long-tail datasets [6,9,15,37,64].”
MethodA dual-branch model with inter- and intra-branch contrastive loss for long-tailed recognition
Chen, Q., Huang, T. S., Zhu, G., & Lin, E. (2023) · Neural Networks · doi:10.1016/j.neunet.2023.09.022 · Cited by 18
“Class re-balancing can be roughly divided into re-sampling (Buda et al., 2018; Pouyanfar et al., 2018; Chawla et al., 2002; Drummond et al., 2003) and re-weighting (Wang et al., 2017; Park et al., 2021; Shu et al., 2019), both methods share the goal of rectifying the skewed training data to improve…”
BackgroundMethodDong, X., Du, W., & Wang, H. F. (2023) · IEEE Transactions on Power Systems · doi:10.1109/tpwrs.2023.3326137 · Cited by 18
“In the real world, class imbalance scenarios generally fall into two main categories: step imbalances and linear imbalances [54].”
BackgroundMachine and Deep Learning Methods for Predicting 3D Genome Organization
Wall, B. P. G., Nguyen, M., Harrell, J. C., & Dozmorov, M. G. (2024) · Methods in molecular biology · doi:10.1007/978-1-0716-4136-1_22 · Cited by 20
Corponi, F., Li, B. M., Anmella, G., Mas, A., Pacchiarotti, I., Valentí, M., Grande, I., Benabarre, A., Garriga, M., Vieta, E., et al. (2024) · Translational Psychiatry · doi:10.1038/s41398-024-02876-1 · Cited by 19
“(ii) Probability thresholding [45]: during inference, probabilistic predictions for each rank under the jt ℎ item were divided by the corresponding rank frequency (computed on the training set), plus a small term to avoid division by zero in case of zero frequency ranks.”
Tuo, J., Zha, M., Li, H., Xie, D., Wang, Y., Sheng, G., & Wang, Y. (2024) · Separation and Purification Technology · doi:10.1016/j.seppur.2024.128237 · Cited by 18
Robust Signal Classification Using Siamese Networks
Langford, Z., Eisenbeiser, L., & Vondal, M. (2019) · ACM Workshop on Wireless Security and Machine Learning (WiseML) · doi:10.1145/3324921.3328781 · Cited by 18
“However, CNNs have several potential issues, such as requiring large amounts of data for training, tending to overfit to particular classes [3], and vulnerability to adversarial examples [6].”
BackgroundSaini, M., & Susan, S. (2019) · Communications in computer and information science · doi:10.1007/978-981-13-9181-1_49 · Cited by 18
Wu, J. F. (2020) · The Astrophysical Journal · doi:10.3847/1538-4357/abacbb · Cited by 18
Group Split and Merge Prediction With 3D Convolutional Networks
Wang, A., & Steinfeld, A. M. (2020) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2020.2969947 · Cited by 18
“However, feeding too much training data from the no-action class leads to the class imbalance problem [2].”
BackgroundKoziarski, M. (2021) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn52387.2021.9533998 · Cited by 20
“[3], who consider the impact of RUS, ROS, and two-phase training, an approach in which the network is first trained on the balanced dataset, and afterwards finetuned on the original data.”
BackgroundA Unified Deep Learning Framework for Short-Duration Speaker Verification in Adverse Environments
Jung, Y., Choi, Y., Lim, H., & Kim, H. (2020) · IEEE Access · doi:10.1109/access.2020.3025941 · Cited by 18
“It has been shown that this class imbalance in training can degrade the performance of deep learningbased classifiers in various domains [62].”
BackgroundAi, X., Sheng, V. S., Fang, W., Ling, C. X., & Li, C. (2020) · IEEE Access · doi:10.1109/access.2020.3035910 · Cited by 18
Fairness-aware Class Imbalanced Learning
Subramanian, S., Rahimi, A., Baldwin, T., Cohn, T., & Frermann, L. (2021) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2021.emnlp-main.155 · Cited by 33
“A skewed class distribution hurts the performance of deep learning models (Buda et al., 2018), and approaches such as instance weighting (Lin et al., 2017; Cui et al., 2019; Li et al., 2020), data augmentation (Juuti et al., 2020; Wei and Zou, 2019), and weighted max-margin (Cao et al., 2019) are…”
BackgroundKim, E., Dahiya, G. S., Løset, S., & Skjetne, R. (2019) · Results in Engineering · doi:10.1016/j.rineng.2019.100036 · Cited by 18
“To improve this uneven class distribution, we used random minority oversampling method proposed by Buda et al. [36].”
MethodELF: An Early-Exiting Framework for Long-Tailed ClassificationPreprint
Duggal, R., Freitas, S., Dhamnani, S., Chau, D. H., & Sun, J. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.11979 · Cited by 29
“On the other hand, undersampling is susceptible to losing information from the majority classes [17].”
BackgroundKapanski, A. A., Hruntovich, N., Klyuev, R. V., Boltrushevich, A. E., Николаевна, С. С., Ефременков, Е. А., Demin, A. Y., & Martyushev, N. V. (2025) · Smart Cities · doi:10.3390/smartcities8020059 · Cited by 18
“The initial statistics then underwent pre-processing, which included removing outliers, filling in missing values, and normalizing features [30].”
Overview of machine learning in class imbalance scenarios: Trends, challenges, and approaches
Hida, G. S., & do Nascimento, A. C. A. (2025) · Expert Systems with Applications · doi:10.1016/j.eswa.2025.129592 · Cited by 23
Computational Landscape in Drug Discovery: From AI/ML Models to Translational Application
Sharma, D., Anabala, M., Jain, V. V., Shyam, M., Prince, S. E., & Muniyan, R. (2025) · Scientifica · doi:10.1155/sci5/1688637 · Cited by 17
Recent Advances in Medical Image Processing
Huang, Z., Li, Q., Ju, L., Feng, J., Hu, J., & Chen, P. (2020) · Acta Cytologica · doi:10.1159/000510992 · Cited by 17
“Many methods have been proposed to address these challenges [92–94].”
MethodAn automatic method for removing empty camera trap images using ensemble learning
Yang, D., Tan, K., Huang, Z., Li, X., Chen, B., Ren, G., & Xiao, W. (2021) · Ecology and Evolution · doi:10.1002/ece3.7591 · Cited by 16
“The performance of the DCNN model is susceptible to the class imbalance of the training set, and the classifier trends to recognize the input as the majority class in the training set (Buda et al., 2018).”
MethodA Probabilistic Formulation for Meta-Weight-Net
Zhao, Q., Shu, J., Yuan, X., Liu, Z., & Meng, D. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3105104 · Cited by 20
“These robust learning issues have also been theoretically investigated in previous work [13]–[17], and become an important while challenging research topic [1], [18].”
BackgroundHandhal, A. M., Ettensohn, F. R., Al-Abadi, A. M., & Ismail, M. J. (2021) · Journal of Petroleum Science and Engineering · doi:10.1016/j.petrol.2021.109482 · Cited by 17
Alleviating ASR Long-Tailed Problem by Decoupling the Learning of Representation and Classification
Deng, K., Cheng, G., Yang, R., & Yan, Y. (2021) · IEEE/ACM Transactions on Audio Speech and Language Processing · doi:10.1109/taslp.2021.3138707 · Cited by 17
Soltani, S., Feilhauer, H., Duker, R., & Kattenborn, T. (2022) · ISPRS Open Journal of Photogrammetry and Remote Sensing · doi:10.1016/j.ophoto.2022.100016 · Cited by 18
LAN Intrusion Detection Using Convolutional Neural Networks
Zainel, H. A., & Koçak, C. (2022) · Applied Sciences · doi:10.3390/app12136645 · Cited by 17
Abibullaev, B., Kunanbayev, K., & Zollanvari, A. (2022) · IEEE Transactions on Human-Machine Systems · doi:10.1109/thms.2022.3189576 · Cited by 18
“Based on the conclusion of Buda et al. [63], we oversample the observations (after preprocessing stage) the minority class to adjust the class distribution balance.”
MethodLuo, W., Zhang, C., Li, Y., Feng, Y., Zhang, D., & Hong, Z. (2022) · International Journal of Applied Earth Observation and Geoinformation · doi:10.1016/j.jag.2022.102949 · Cited by 18
Adaptive cost-sensitive stance classification model for rumor detection in social networks
Zojaji, Z., & Ladani, B. T. (2022) · Social Network Analysis and Mining · doi:10.1007/s13278-022-00952-2 · Cited by 17
“Finally, random undersampling, random oversampling, and two-phase learning were applied and compared over several imbalanced image datasets in (Buda et al. 2018).”
MethodTowards Automated Imbalanced Learning with Deep Hierarchical Reinforcement Learning
Zha, D., Lai, K.-H., Tan, Q., Ding, S., Zou, N., & Hu, X. (2022) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3511808.3557474 · Cited by 17
“Since these datasets are relatively balanced, following the previous work [2, 38, 54], we artificially create the imbalanced datasets by randomly under-sampling the minority class with different imbalanced ratios.”
BackgroundEnd to End Multitask Joint Learning Model for Osteoporosis Classification in CT Images
Zhang, K., Lin, P., Pan, J., Pei-xia, X., Qiu, X., Crookes, D., Hua, L., & Wang, L. (2023) · Computational Intelligence and Neuroscience · doi:10.1155/2023/3018320 · Cited by 22
“At present, the methods to solve the data imbalance include data resampling [33], adaptive loss function [34], and curriculum learning [35].”
MethodUnderstanding the Challenges When 3D Semantic Segmentation Faces Class Imbalanced and OOD Data
Pan, Y., Xie, F., & Zhao, H. (2023) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2023.3256442 · Cited by 17
“Although many methods [10], [33] have Fig.”
BackgroundInfluentialD’Arco, L., Wang, H., & Zheng, H. (2023) · Neural Computing and Applications · doi:10.1007/s00521-023-08363-w · Cited by 17
“In general, two strategies could be applied [51]: re-sampling [52, 53] and cost-sensitive re-weighting [54, 55].”
MethodLinear-exponential loss incorporated deep learning for imbalanced classification
Fu, S., Su, D., Li, S., Sun, S., & Tian, Y. (2023) · ISA Transactions · doi:10.1016/j.isatra.2023.06.016 · Cited by 18
Hierarchical-Instance Contrastive Learning for Minority Detection on Imbalanced Medical Datasets
Li, Y., Qian, G., Jiang, X., Jiang, Z., Wen, W., Zhang, S., Li, K., & Lao, Q. (2023) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2023.3310716 · Cited by 18
“Re-sampling use over-sampling [4], [23], [24] to increase the number of samples in the minority classes or under-sampling [5], [25], [26] by removing samples for majority classes, which often leads to either over-fitting or under-fitting.”
BackgroundMethodAugmentation strategies for an imbalanced learning problem on a novel COVID-19 severity dataset
Schaudt, D., von Schwerin, R., Hafner, A., Riedel, P., Reichert, M., von Schwerin, M., Beer, M., & Kloth, C. (2023) · Scientific Reports · doi:10.1038/s41598-023-45532-2 · Cited by 23
“Therefore, ROS is prefered in most cases 48 .”
BackgroundLee, S.-J., Oh, H. J., Son, Y., Kim, J., Kwon, I.-J., Kim, B., Lee, J.-H., & Kim, H.-K. (2024) · BMC Oral Health · doi:10.1186/s12903-024-03898-3 · Cited by 21
“This disparity can degrade classifier performance by neglecting minority classes [21–24].”
BackgroundLong-Tailed Anomaly Detection with Learnable Class Names
Ho, C.-H., Peng, K., & Vasconcelos, N. M. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.01182 · Cited by 21
“Data re-sampling meth-ods [6, 13, 15, 22, 27, 28, 75] balance the sample distribution by under-sampling majority classes or augmenting minority classes.”
BackgroundSemantic segmentation on small datasets of satellite images using convolutional neural networks
Younis, M. C., & Keedwell, E. C. (2019) · Journal of Applied Remote Sensing · doi:10.1117/1.jrs.13.046510 · Cited by 17
“As the class distribution in our dataset was not balanced (the number of pixels related to vegetation was higher than the other classes), the SoftMax computation assigns a weight to each class, 43 based on the inverse of the frequency of occurrence (the rarer the class, the higher the weight).”
BackgroundCell Nuclei Classification in Histopathological Images using Hybrid O L ConvNet
Tripathi, S., & Singh, S. K. (2020) · ACM Transactions on Multimedia Computing Communications and Applications · doi:10.1145/3345318 · Cited by 17
“A systematic study of the class imbalance problem in convolutional neural networks by authors in [41] have concluded through their research that the class imbalance problem, if not addressed may have a detrimental effect on classification performance.”
BackgroundWang, X. R., Jin, Y., & Zhang, Q. (2020) · International Journal of Molecular Sciences · doi:10.3390/ijms21165710 · Cited by 17
“Over-sampling is commonly used in deep learning [32,33].”
BackgroundPeng, D., Gu, T., Hu, X. L., & Liu, C. (2021) · Neurocomputing · doi:10.1016/j.neucom.2020.12.122 · Cited by 17
Minority Class Oriented Active Learning for Imbalanced Datasets
Aggarwal, U., Popescu, A., & Hudelot, C. (2021) · Proceedings - International Conference on Pattern Recognition/Proceedings/International Conference on Pattern Recognition · doi:10.1109/icpr48806.2021.9412182 · Cited by 17
“However, in practice, data are often unevenly distributed across classes and imbalance needs to be dealt with [9].”
BackgroundInfluentialIB-GAN: A Unified Approach for Multivariate Time Series Classification under Class Imbalance
Deng, G., Han, C., Dreossi, T., Lee, C. C., & Matteson, D. S. (2022) · Society for Industrial and Applied Mathematics eBooks · doi:10.1137/1.9781611977172.25 · Cited by 17
“The three baseline methods for addressing imbalances are class weights, upsampling, and downsampling, each with drawbacks [6, 7].”
MethodClustering and Learning from Imbalanced DataPreprint
Singh, N. D., & Dhall, A. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1811.00972 · Cited by 24
“Authors of [5] study the effects of imbalance on convolutional neural networks and conclude that oversampling is the best-suited resampling technique and does not lead to overfitting in CNNs.”
BackgroundYu, H., Zhang, N., Deng, S., Yuan, Z., Jia, Y., & Chen, H. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2009.07022 · Cited by 28
“There are two main methods by which the long-tailed issue can be addressed (Buda et al., 2018).”
Contrastive Learning based Hybrid Networks for Long-Tailed Image ClassificationPreprint
Wang, P., Han, K., Wei, X.-S., Lei Zhang, A., & Wang, L. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.14267 · Cited by 17
“Typical methods include data re-sampling [1, 8, 35, 4, 28], loss re-weighting [26, 7, 30, 33], margin modification [3], and data augmentation [19, 6, 29].”
BackgroundMethodtanh As a robust feature scaling method in training deep learning models with imbalanced data
Yang, A., Zhang, M., Chen, H., Li, T., Liu, S., & Xu, X. (2025) · Pattern Recognition · doi:10.1016/j.patcog.2025.111746 · Cited by 16
Zhou, Y., Jiang, D., Wei, X. L., Yi, J., Wang, Y., Deng, Y., & Cao, D. (2025) · Chemical Science · doi:10.1039/d5sc04631a · Cited by 16
“In the training of multi-task models, loss weighting strategies 37 have been introduced to address the imbalance between positive and negative samples in the dataset, enabling the model to pay more attention to underrepresented classes during training.”
ROSNet: Robust one-stage network for CT lesion detection
Lung, K.-Y., Chang, C.-R., Weng, S.-E., Lin, H.-S., Shuai, H.-H., & Cheng, W.-H. (2021) · Pattern Recognition Letters · doi:10.1016/j.patrec.2021.01.011 · Cited by 17
Gao, X., & Mathur, S. (2021) · Journal of Climate · doi:10.1175/jcli-d-21-0137.1 · Cited by 16
Ghayvat, H., & Gope, P. (2021) · Neural Computing and Applications · doi:10.1007/s00521-021-06139-8 · Cited by 16
Reliable identification and quantification of neural cells in microscopic images of neurospheres
Förster, N., Butke, J., Keßel, H. E., Bendt, F., Pahl, M., Li, L., Fan, X., Leung, P., Klose, J., Masjosthusmann, S., et al. (2021) · Cytometry Part A · doi:10.1002/cyto.a.24514 · Cited by 16
“Class imbalance is a notorious problem when training and validating classification models and has thus been investigated in much detail, in particular in the context of training convolutional neural networks [37, 38].”
BackgroundAugmenting data with generative adversarial networks: An overview
Ljubić, H., Martinović, G., & Volarić, T. (2022) · Intelligent Data Analysis · doi:10.3233/ida-215735 · Cited by 16
“Generally, algorithm modifications can be divided into 4 different approaches: (i) thresholding; (ii) one-class classification (OCC); (iii) cost-sensitive learning; and (iv) hybrid model [16].”
A Deep Dive into Dataset Imbalance and Bias in Face Identification
Cherepanova, V., Reich, S., Dooley, S. W., Souri, H., Dickerson, J., Goldblum, M., & Goldstein, T. (2023) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1145/3600211.3604691 · Cited by 16
“We refer the reader to [3, 21], for a thorough review of deep learningbased imbalance literature.”
BackgroundTowards prior gap and representation gap for long-tailed recognition
Zhang, M.-L., Zhang, X.-Y., Wang, C., & Liu, C. (2022) · Pattern Recognition · doi:10.1016/j.patcog.2022.109012 · Cited by 16
Tongue Segmentation and Color Classification Using Deep Convolutional Neural Networks
Yan, B., Zhang, S., Yang, Z., Su, H.-Y., & Zheng, H. (2022) · Mathematics · doi:10.3390/math10224286 · Cited by 16
Aligning model outputs for class imbalanced non-IID federated learning
Li, L., Zhan, D., & Li, X. (2022) · Machine Learning · doi:10.1007/s10994-022-06241-5 · Cited by 20
“While global data balance is a strong assumption, the data often follows a long-tailed distribution in real-world scenarios (Buda et al., 2018).”
An adaptive machine learning algorithm for the resource-constrained classification problem
Shifman, D. A., Cohen, I., Huang, K., Xian, X., & Singer, G. (2022) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2022.105741 · Cited by 18
“Many studies have applied cost-sensitive approaches to handling imbalanced classification problems [13, 14] where the decision maker is interested in detecting the positive cases.”
BackgroundGuo, X., Guo, W., Liu, J., Qiao, J., & Hu, G. (2023) · Underground Space · doi:10.1016/j.undsp.2023.01.004 · Cited by 16
Recognition of Edible Fungi Fruit Body Diseases Based on Improved ShuffleNetV2
Xu, X., Zhang, Y., Cao, H., Yang, D., Zhou, L., & Yu, H. (2023) · Agronomy · doi:10.3390/agronomy13061530 · Cited by 16
“The balance of the test data is ensured by data augmentation [50], the quality and quantity of data samples are improved, and sufficient data samples are provided for training the convolutional neural network [51].”
Gao, Z., Wang, J., Yu, G., Yan, Z., Domeniconi, C., & Zhang, J. (2023) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v37i6.25927 · Cited by 16
Probability Guided Loss for Long-Tailed Multi-Label Image Classification
Lin, D. (2023) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v37i2.25244 · Cited by 19
“The basic idea is to over-sample the minority categories or to under-sample the frequent categories during training (Han, Wang, and Mao 2005; Buda, Maki, and Mazurowski 2018).”
BackgroundDynamic Loss for Robust Learning
Jiang, S., Li, J., Zhang, J., Wang, Y., & Xu, T. (2023) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2023.3311636 · Cited by 16
“Previous works on long-tailed learning can be broadly categorized into three main approaches: data re-sampling [20], [21], boundary adjustment [22], and re-weighting [9], [23], [24].”
BackgroundTowards Long-Tailed Recognition for Graph Classification via Collaborative Experts
Yi, S., Mao, Z., Ju, W., Zhou, Y., Liu, L., Luo, X., & Zhang, M. (2023) · IEEE Transactions on Big Data · doi:10.1109/tbdata.2023.3313029 · Cited by 16
“…learning aims to alleviate the impact of class imbalance on model training, and there are currently three main strategies to address this practical problem: re-sampling [15], [16], [34], [35], [36], [37], re-weighting [22], [38], [39], [40], [41], and ensemble learning [18], [26], [42], [43], [44].”
BackgroundZhang, K., Zhang, R., Yang, Z., Deng, J., Abdullah, A., Zhou, C., Lv, X., Wang, R., & Ma, Z. (2023) · Remote Sensing · doi:10.3390/rs15184572 · Cited by 16
Feng, Z., Zhou, R., Xia, W., Wang, S., Liu, Y., Huang, Z., & Gan, H. (2023) · Medical Physics · doi:10.1002/mp.16875 · Cited by 16
“39 Balanced MSE is suggested as the loss function for training and evaluating the model to mitigate the impact of imbalanced data distribution on model performance.”
Gill, T. S., Zaidi, S. S. H., & Shirazi, M. A. (2024) · Epilepsy & Behavior · doi:10.1016/j.yebeh.2024.109732 · Cited by 16
Sun, G., Zhang, Y., Wang, L., Zhou, L., Fei, S., Han, S., Xiao, S., Che, Y., Yan, L., Xu, Y., et al. (2025) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2024.109859 · Cited by 16
Koziarski, M., Kwolek, B., & Cyganek, B. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-12177-8_1 · Cited by 16
Valid oversampling schemes to handle imbalance
Kim, Y.-G., Kwon, Y., & Paik, M. C. (2019) · Pattern Recognition Letters · doi:10.1016/j.patrec.2019.07.006 · Cited by 16
Efficient Incremental Learning Using Dynamic Correction Vector
Xiang, Y., Miao, Y., Chen, J., & Xuan, Q. (2020) · IEEE Access · doi:10.1109/access.2019.2963461 · Cited by 16
Rivas, R., Sadah, S. A., Guo, Y., & Hristidis, V. (2020) · JMIR Public Health and Surveillance · doi:10.2196/14952 · Cited by 20
“These weights are used with random forest and SVM according to their implementations by Pedregosa et al [36], whereas CNN uses oversampling of the least frequent class as recommended by Buda et al [37].”
MethodFˆ2-Softmax: Diversifying Neural Text Generation via Frequency Factorized Softmax
Choi, B.-J., Hong, J., Park, D., & Lee, S. W. (2020) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.737 · Cited by 16
“One of the widely adopted techniques for imbalanced classification is re-sampling, which includes removing examples from the majority classes (under-sampling) and adding samples for the minority classes (over-sampling) (Buda et al., 2018).”
BackgroundMethodInfluentialTowards scaling Twitter for digital epidemiology of birth defects
Klein, A. Z., Sarker, A., Weissenbacher, D., & Gonzalez‐Hernandez, G. (2019) · npj Digital Medicine · doi:10.1038/s41746-019-0170-5 · Cited by 16
“ased classifiers are generally outperformed for imbalanced social media data [46]. Our results also suggest that datalevel approaches for addressing class imbalance with convolutional neural networks [47] may not generalize to recurrent neural networks (e.g., LSTMs). 9 The F 1 -scores for the minority classes of the S…”
BackgroundInfluentialRudmann, D. G., Albretsen, J. C., Doolan, C., Gregson, M., Dray, B. K., Sargeant, A. M., O'Shea, D. G., Kuklyte, J., Power, A., & Fitzgerald, J. (2020) · Toxicologic Pathology · doi:10.1177/0192623320973986 · Cited by 16
“33 In this work, a comparably small number of annotations (313-1018 tiles per organ) and tissue sections (17-35 per organ) were used, and the proportion of lesion pixels in each annotation set varied from 3.5 % (stomach lesions at (cid”
BackgroundMethodInfluentialInvestigating the Homogenization of Web Design: A Mixed-Methods Approach
Goree, S., Doosti, B., Crandall, D., & Su, N. M. (2021) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3411764.3445156 · Cited by 17
“In addition to the reasons above, we use the Dense Russell instead of the full Russell dataset because CNN models are known to be sensitive to class imbalance [16], so it is important that each of the classes has a roughly equal number of snapshots.”
MethodImbalanced Adversarial Training with Reweighting
Wang, W., Xu, H., Liu, X., Li, Y., Thuraisingham, B. M., & Tang, J. (2022) · Proceedings · doi:10.1109/icdm54844.2022.00156 · Cited by 20
“Specifically, we consider the Step imbalance [20] setting following the similar process adopted in Section II-A.”
BackgroundMethodTsai, D., Huang, Y.-Q., & Chiu, W. (2021) · Measurement Science and Technology · doi:10.1088/1361-6501/ac1fbf · Cited by 17
Noor, N. S. E. M., Ibrahim, H., Lah, M. H. C., & Abdullah, J. M. (2021) · IEEE Access · doi:10.1109/access.2021.3109780 · Cited by 18
Metadata
- Authors
- Buda, M., Maki, A., & Mazurowski, M. A.
- Affiliations
- Duke University (US)
- KTH Royal Institute of Technology (SE)
- Corresponding author
- Buda, Mateusz
- Venue
- Neural Networks
- Publisher
- Elsevier BV
- Published
- 2018-10
- Volume
- 106
- Pages
- 249-259
- ISSN
- 0893-6080 (ISSN-L)
- OpenAlex
- W2767106145
- Cited by
- 3,088
- References
- 83
- Citations by year
- 2026: 459 · 2025: 507 · 2024: 371 · 2023: 405 · 2022: 405 · 2021: 380 · 2020: 308 · 2019: 189 · 2018: 58
- FWCI
- 131.58
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Neural Networks (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- KTH Publication Database DiVA (KTH Royal Institute of Technology) (submittedVersion, other-oa)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Imbalanced Data Classification Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Keywords
- class imbalance, convolutional neural networks, oversampling, undersampling, two-stage training, thresholding, area under the receiver operating characteristic, multi-class classification, overfitting, MNIST, CIFAR-10, ImageNet
- MeSH terms
- Machine Learning, Humans, Probability, ROC Curve, Neural Networks, Computer
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
