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 401–500 of 1,001
A solanaceae disease recognition model based on SE-Inception
Li, Z., Yang, Y., Li, Y., Guo, R., Yang, J., & Yue, J. (2020) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2020.105792 · Cited by 50
Sugino, T., Kawase, T., Onogi, S., Kin, T., Saito, N., & Nakajima, Y. (2021) · Healthcare · doi:10.3390/healthcare9080938 · Cited by 50
“To address this problem, which is commonly known as the class imbalance, there are two types of approaches: data-level approaches and algorithm-level approaches [7,8].”
MethodPredicting Usual Interstitial Pneumonia Histopathology From Chest CT Imaging With Deep Learning
Bratt, A. K., Williams, J. M., Liu, G., Panda, A., Patel, P. P., Walkoff, L. A., Sykes, A.-M. G., Tandon, Y. K., François, C. J., Blezek, D. J., et al. (2022) · CHEST Journal · doi:10.1016/j.chest.2022.03.044 · Cited by 49
Synthetic Data Generation to Speed-Up the Object Recognition Pipeline
Perri, D., Simonetti, M., & Gervasi, O. (2021) · Electronics · doi:10.3390/electronics11010002 · Cited by 50
PD-ResNet for Classification of Parkinson’s Disease From Gait
Yang, X., Ye, Q., Cai, G., Wang, Y.-Q., & Cai, G. (2022) · IEEE Journal of Translational Engineering in Health and Medicine · doi:10.1109/jtehm.2022.3180933 · Cited by 57
“Random minority oversam-pling methods replicate samples from minority classes, which may cause overfitting [10], [24].”
Hou, C., Zhang, X., Tang, Y., Zhuang, J., Tan, Z., Huang, H., Chen, W., Wei, S., He, Y., & Luo, S. (2022) · Frontiers in Plant Science · doi:10.3389/fpls.2022.972445 · Cited by 49
“Notably, the target-background class imbalance is typically the main obstacle encountered in training convolutional neural networks (Buda et al., 2018).”
BackgroundA Multidepth and Multibranch Network for Hyperspectral Target Detection Based on Band Selection
Gao, H., Zhang, Y., Chen, Z., Xu, S., Hong, D., & Zhang, B. (2023) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2023.3258061 · Cited by 53
Subclass-balancing Contrastive Learning for Long-tailed Recognition
Hou, C., Zhang, J., Wang, H., & Zhou, T. (2023) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00497 · Cited by 50
“However, training on such class-imbalanced datasets can result in a severely biased model with noticeable performance drop in classification tasks [2,4,13,47,64,68, 71].”
BackgroundMethodAhamed, M. F., Shafi, F. B., Nahiduzzaman, M., Ayari, M. A., & Khandakar, A. A. (2024) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2024.109503 · Cited by 49
Deep hashing for multi-label image retrieval: a survey
Rodrigues, J., de Cristo, M. A. P., & Colonna, J. G. (2020) · Artificial Intelligence Review · doi:10.1007/s10462-020-09820-x · Cited by 50
“…image retrieval methods are: Normalized Discounted Cumulative Gains (NDCG) (Järve-lin and Kekäläinen 2002), Mean Average Precision (MAP) (Baeza-Yates and Ribeiro-Neto 1999), Weighted MAP (Baeza-Yates and Ribeiro-Neto 1999) and Average Cumulative Gains (ACG) (Järvelin and Kekäläinen 2000).”
MethodInfluentialPolarimetric Multipath Convolutional Neural Network for PolSAR Image Classification
Cui, Y., Liu, F., Jiao, L., Guo, Y., Liang, X., Li, L., Yang, S., & Qian, X. (2021) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2021.3071559 · Cited by 51
An ensemble of deep transfer learning models for handwritten music symbol recognition
Paul, A., Pramanik, R., Malakar, S., & Sarkar, R. (2021) · Neural Computing and Applications · doi:10.1007/s00521-021-06629-9 · Cited by 48
“It is to be noted that these datasets suffer from class imbalance problems and traditionally deep learning models are known to handle the class imbalance problem quite well [38].”
BackgroundBalancing Logit Variation for Long-Tailed Semantic Segmentation
Wang, Y., Fei, J., Wang, H., Li, W., Bao, T., Wu, L., Zhao, R., & Shen, Y. (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/cvpr52729.2023.01874 · Cited by 54
“One intuitive solution to alleviate unbalanced data distribution is data processing, which typically consists of three ways: over-sampling [34,35,68,89], under-sampling [6,35, 76, 80] and data augmentation [15, 16, 56, 104].”
BackgroundDeep Learning and Data Sampling with Imbalanced Big Data
Johnson, J. M., & Khoshgoftaar, T. M. (2019) · IEEE International Conference on Information Reuse and Integration (IRI) · doi:10.1109/iri.2019.00038 · Cited by 55
“[39] presented similar results when they compared ROS with RUS.”
BackgroundResultCross-Domain Empirical Risk Minimization for Unbiased Long-Tailed Classification
Zhu, B., Niu, Y., Hua, X., & Zhang, H. (2022) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v36i3.20271 · Cited by 49
“…the contribution of each class in the training phase, which can be divided into two strategies: re-sampling the data to balance the class frequency (Buda, Maki, and Mazurowski 2018; Drumnond 2003; Shen, Lin, and Huang 2016; Sarafi-anos, Xu, and Kakadiaris 2018; Japkowicz and Stephen 2002), and…”
MethodCSCNN: Cost-Sensitive Convolutional Neural Network for Encrypted Traffic Classification
Soleymanpour, S., Sadr, H., & Soleimandarabi, M. N. (2021) · Neural Processing Letters · doi:10.1007/s11063-021-10534-6 · Cited by 47
Han, Z., Hu, W., Peng, S., Lin, H., Zhang, J., Zhou, J., Wang, P., & Dian, Y. (2022) · Remote Sensing · doi:10.3390/rs14133075 · Cited by 47
“Moreover, the imbalance and small sample size between the target and the background are still the main factors limiting the detection accuracy [27].”
BackgroundDeep active learning models for imbalanced image classification
Jin, Q., Yuan, M., Wang, H., Wang, M., & Song, Z. (2022) · Knowledge-Based Systems · doi:10.1016/j.knosys.2022.109817 · Cited by 50
Zhang, Y., Zhan, Q., & Ma, Z. (2024) · Advanced Engineering Informatics · doi:10.1016/j.aei.2024.102737 · Cited by 47
Prototyping a Social Media Flooding Photo Screening System Based on Deep Learning
Ning, H., Li, Z., Hodgson, M. E., & Wang, C. (2020) · ISPRS International Journal of Geo-Information · doi:10.3390/ijgi9020104 · Cited by 47
“The plug-in-based design of RIASM makes it extendable for supporting other types of disaster events such as wildfires and earthquakes for the damage / impact assessment as well as other studies beyond disaster management.”
MethodPSNet: A Style Transfer Network for Point Cloud Stylization on Geometry and Color
Cao, X., Wang, W., Nagao, K., & Nakamura, R. (2020) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv45572.2020.9093513 · Cited by 47
“a combination of undersampling and oversampling techniques [2]: we undersample or oversample 320 instances for classes in which instances are over or under 320, respectively.”
MethodBarata, F., Tinschert, P., Rassouli, F., Steurer-Stey, C. A., Fleisch, E., Puhan, M. A., Brutsche, M. H., Kotz, D. F., & Kowatsch, T. (2020) · Journal of Medical Internet Research · doi:10.2196/18082 · Cited by 46
“This class imbalance can have a considerable detrimental effect on convergence during the training phase and generalization of a model on the test set [40].”
BackgroundMethodHu, K., Huang, Y., Huang, W., Tan, H. Y., Chen, Z., Zheng, Z., Li, X., Zhang, Y., & Gao, X. (2021) · Neurocomputing · doi:10.1016/j.neucom.2021.06.012 · Cited by 47
“, data-level methods and algorithm-level methods [5].”
BackgroundGao, X., & Wang, X. (2019) · Diagnostic and Interventional Imaging · doi:10.1016/j.diii.2019.07.002 · Cited by 47
Memory-Based Jitter: Improving Visual Recognition on Long-Tailed Data with Diversity in Memory
Liu, J., Li, W., & Sun, Y. (2022) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v36i2.20064 · Cited by 49
“Such long-tailed distribution significantly challenges deep visual recognition, including both deep image classification [48, 12, 4, 35, 1, 6, 50] and deep metric learning [17, 42, 8, 32].”
BackgroundGaNDLF: the generally nuanced deep learning framework for scalable end-to-end clinical workflows
Pati, S., Thakur, S. P., Hamamcı, İ. E., Baid, U., Baheti, B., Bhalerao, M., Güley, O., Mouchtaris, S., Lang, D., Thermos, S., et al. (2023) · Communications Engineering · doi:10.1038/s44172-023-00066-3 · Cited by 68
Usman, O. L., & Muniyandi, R. C. (2020) · Symmetry · doi:10.3390/sym12050836 · Cited by 50
“ROC and Area under ROC (AUC): The receiver operating characteristics (ROC) curve plots the sensitivity curve against specificity, and thus provides a representation of the trade-off between correctly classified positive instances and incorrectly classified negative instances [83].”
BackgroundFew‐shot learning for plant disease recognition: A review
Sun, J., Cao, W., Xi, F., Ochi, S., & Yamanaka, T. (2023) · Agronomy Journal · doi:10.1002/agj2.21285 · Cited by 49
A survey of class-imbalanced semi-supervised learning
Gui, Q., Hong, Z., Guo, N., & Niu, B. (2023) · Machine Learning · doi:10.1007/s10994-023-06344-7 · Cited by 51
“…where the number of data points exponential decline from the largest class to the smallest class, i.e. n k = n 1 ∗ 1 − k L − 1 , and the step imbalance (Buda et al., 2018) where the majority classes have same number of data points and the minority classes also have the same number of data points.”
InfluentialBalanced Product of Calibrated Experts for Long-Tailed Recognition
Aimar, E. S., Jonnarth, A., Felsberg, M., & Kuhlmann, M. (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/cvpr52729.2023.01912 · Cited by 48
“ance by different means, including data re-sampling [4, 8], cost-sensitive learning [1, 65], and margin modifications [6, 56].”
BackgroundYue, J., Wang, J., Zhang, Z., Li, C., Yang, H. F., Feng, H., & Guo, W. F. (2024) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2024.109653 · Cited by 48
DeepNeuro: an open-source deep learning toolbox for neuroimaging
Beers, A. L., Brown, J. M., Chang, K., Hoebel, K., Patel, J., Ly, K. I., Tolaney, S. M., Brastianos, P. K., Rosen, B. R., Gerstner, E. R., et al. (2020) · Neuroinformatics · doi:10.1007/s12021-020-09477-5 · Cited by 49
Anchor Loss: Modulating Loss Scale Based on Prediction Difficulty
Ryou, S., Jeong, S.-G., & Perona, P. (2019) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00609 · Cited by 44
“Typical strategies to mitigate this issue are class re-sampling [8, 18, 6] or cost-sensitive learning [50, 23, 14].”
BackgroundMethodFew-shot learning for dermatological disease diagnosis
Prabhu, V., Kannan, A., Ravuri, M., Chablani, M., Sontag, D. A., & Amatriain, X. (2022) · Elsevier eBooks · doi:10.1016/b978-0-32-399851-2.00022-3 · Cited by 46
Kim, M., Ou, E., Loh, P., Allen, T. R., Agasie, R. J., & Liu, K. (2020) · Nuclear Engineering and Design · doi:10.1016/j.nucengdes.2020.110699 · Cited by 45
Chen, S.-S., Choubey, B., & Singh, V. K. (2021) · Journal of Retailing and Consumer Services · doi:10.1016/j.jretconser.2021.102573 · Cited by 50
Ryu, J. Y., Lee, J. H., Lee, B. H., Song, J. S., Ahn, S., & Oh, K. (2021) · Bioinformatics · doi:10.1093/bioinformatics/btab547 · Cited by 45
Mehrvar, S., Himmel, L. E., Babburi, P., Goldberg, A. L., Guffroy, M., Janardhan, K. S., Krempley, A. L., & Bawa, B. (2021) · Journal of Pathology Informatics · doi:10.4103/jpi.jpi_36_21 · Cited by 46
Graph Learning-Based Arithmetic Block Identification
He, Z., Wang, Z., Bai, C., Yang, H., & Yu, B. (2021) · doi:10.1109/iccad51958.2021.9643581 · Cited by 44
“We adopt the basic version of it, called random minority oversampling, which simply replicates randomly selected samples from minority classes [30].”
BackgroundMethodAutomated Scoring of Respiratory Events in Sleep With a Single Effort Belt and Deep Neural Networks
Nassi, T. E., Ganglberger, W., Sun, H., Bucklin, A. A., Biswal, S., van Putten, M. J. A. M., THOMAS, R. J., & Westover, M. B. (2021) · IEEE Transactions on Biomedical Engineering · doi:10.1109/tbme.2021.3136753 · Cited by 44
“Classification with imbalanced data is challenging in many real-world deep learning applications [32], [33].”
BackgroundBWGAN-GP: An EEG Data Generation Method for Class Imbalance Problem in RSVP Tasks
Xu, M., Chen, Y., Wang, Y., Wang, D., Liu, Z., & Zhang, L. (2022) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2022.3145515 · Cited by 44
Long-tailed visual recognition with deep models: A methodological survey and evaluation
Fu, Y., Xiang, L., Zahid, Y., Ding, G., Mei, T., Shen, Q., & Han, J. (2022) · Neurocomputing · doi:10.1016/j.neucom.2022.08.031 · Cited by 47
Man, C. K., Quddus, M., Theofilatos, A., Yu, R., & Imprialou, M. (2022) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2022.3207798 · Cited by 45
“Sampling is the most straightforward method in handling an imbalanced dataset [53].”
MethodSoftware Defect Prediction Approach Based on a Diversity Ensemble Combined With Neural Network
Chen, J., Xu, J., Cai, S., Wang, X., Chen, H., & Li, Z. (2024) · IEEE Transactions on Reliability · doi:10.1109/tr.2024.3356515 · Cited by 44
Handling imbalance in an extended PLAID
Baets, L. D., Develder, C., Dhaene, T., Deschrijver, D., Gao, J.-K., & Berges, M. (2017) · Sustainable Internet and ICT for Sustainability · doi:10.23919/sustainit.2017.8379795 · Cited by 42
Remote Sensing Big Data Classification with High Performance Distributed Deep Learning
Sedona, R., Cavallaro, G., Jitsev, J., Strube, A., Riedel, M., & Benediktsson, J. A. (2019) · Remote Sensing · doi:10.3390/rs11243056 · Cited by 47
“However, the problem of class imbalance may require the use of of different techniques, e.g., upsampling of the under-represented samples [65] or loss weighting to let the model give more importance to samples associated with classes present in a lesser amount [64].”
MethodCollaborative Deep Learning Models to Handle Class Imbalance in FlowCam Plankton Imagery
Kerr, T., Clark, J. R., Fileman, E. S., Widdicombe, C. E., & Pugeault, N. (2020) · IEEE Access · doi:10.1109/access.2020.3022242 · Cited by 48
“Dealing with imbalanced datasets is a long-standing research topic in machine learning [28], and within ConvNets [41].”
BackgroundGAN-Based Anomaly Detection In Imbalance Problems
Kim, J., Jeong, K., Choi, H., & Seo, K. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-65414-6_11 · Cited by 44
Deep-Feature-Based Autoencoder Network for Few-Shot Malicious Traffic Detection
He, M., Wang, X., Zhou, J., Xi, Y., Jin, L., & Wang, X. (2021) · Security and Communication Networks · doi:10.1155/2021/6659022 · Cited by 44
Lomonaco, V., Pellegrini, L., Rodríguez, P., Caccia, M., She, Q., Chen, Y., Jodelet, Q., Wang, R., Mai, Z., Vázquez, D., et al. (2021) · Artificial Intelligence · doi:10.1016/j.artint.2021.103635 · Cited by 50
“For NIC track, based on the assumption that the neural network estimates Bayesian a posteriori probilitities (Richard & Lippmann, 1991), the network outputs are divided by the prior probability for each class inspired by the trick that handles class imbalance (Buda et al., 2018).”
MethodA Hybrid Deep Learning Approach for Intrusion Detection in IoT Networks
Emeç, M., & Özcanhan, M. H. (2022) · Advances in Electrical and Computer Engineering · doi:10.4316/aece.2022.01001 · Cited by 46
Zhou, Z., Zheng, C., Liu, X., Tian, Y., Chen, X., Chen, X., & Dong, Z. (2023) · Remote Sensing · doi:10.3390/rs15071768 · Cited by 45
Taha, Z. K., Yaw, C. T., Koh, S. P., Kiong, T. S., Kadirgama, K., Benedict, F., Tan, J. D., & Balasubramaniam, Y. (2023) · IEEE Access · doi:10.1109/access.2023.3267964 · Cited by 48
“This assumption is not true because, in real-world situations, the data frequently exhibits a long-tailed distribution [21].”
BackgroundSharma, O., Sharma, A., & Kalia, A. (2023) · Expert Systems with Applications · doi:10.1016/j.eswa.2023.122678 · Cited by 46
Tracing the bias loop: AI, cultural heritage and bias-mitigating in practice
Foka, A., Griffin, G., Ortiz Pablo, D., Rajkowska, P., & Badri, S. (2025) · AI & SOCIETY · doi:10.1007/s00146-025-02349-z · Cited by 43
“1 Building on the approach by Buda et al. (2018), we applied these augmentation techniques to the sub-classes in both training datasets.”
Visual Question Answering as a Meta Learning Task
Teney, D., & van den Hengel, A. (2018) · Lecture notes in computer science · doi:10.1007/978-3-030-01267-0_14 · Cited by 44
“To provide the baseline with a fair chance2, we train all models with standard supersampling [9, 16], i.”
MethodSolving Long-Tailed Recognition with Deep Realistic Taxonomic Classifier
Wu, T.-Y., Morgado, P. M., Wang, P., Ho, C.-H., & Vasconcelos, N. M. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-58598-3_11 · Cited by 45
“Popular attempts to overcome this limitation include data resampling [6,8,20,30], cost-sensitive losses [14], knowledge transfer from high to low population classes [46,60], normalization [38], or margin-based methods [7].”
BackgroundYoo, J., Jun, T. J., & Kim, Y. (2021) · Computer Methods and Programs in Biomedicine · doi:10.1016/j.cmpb.2021.106281 · Cited by 43
LCReg: Long-tailed image classification with Latent Categories based Recognition
Liu, W., Wu, Z., Wang, Y., Ding, H., Liu, F., Lin, J., & Lin, G. (2023) · Pattern Recognition · doi:10.1016/j.patcog.2023.109971 · Cited by 43
“Random under-sampling involves randomly deleting instances from the more represented classes (also known as ‘head classes’) until they have the same number of instances as the less represented classes [11].”
BackgroundMethodGAT-COBO: Cost-Sensitive Graph Neural Network for Telecom Fraud Detection
Hu, X., Chen, H.-T., Zhang, J., Chen, H.-C., Liu, S.-X., Li, X., Wang, Y.-H., & Xue, X.-Y. (2023) · IEEE Transactions on Big Data · doi:10.1109/tbdata.2024.3352978 · Cited by 41
“The works [7], [16], [52], [53] also provided a review of solutions to the class imbalance problem.”
BackgroundDeep Learning Transfer with AlexNet for chest X-ray COVID-19 recognition
Pérez, E. C., & Sánchez-Sánchez, S. (2021) · IEEE Latin America Transactions · doi:10.1109/tla.2021.9451239 · Cited by 44
“…formas, particularmente en el contexto de clasificación binaria (COVID-19 vs normal), comúnmente algunas de las clases pueden estar sobre-representadas o bajo-representadas, para esto, se emplean dos métricas de desequilibrio [36], la primera es denominada step imbalance, la cual se define en (1).”
BackgroundA Novel Focal Phi Loss for Power Line Segmentation with Auxiliary Classifier U-Net
Jaffari, R., Hashmani, M. A., & Reyes‐Aldasoro, C. C. (2021) · Sensors · doi:10.3390/s21082803 · Cited by 42
“This class imbalance introduces a high bias towards the majority class (background in this case), making it difficult for the deep networks to classify the foreground class of interest, also known as the minority class (PL in this case) accurately [9].”
BackgroundArSL-CNN a convolutional neural network for Arabic sign language gesture recognition
Alani, A. A., & Cosma, G. (2021) · Indonesian Journal of Electrical Engineering and Computer Science · doi:10.11591/ijeecs.v22.i2.pp1096-1107 · Cited by 42
“To overcome this issue and reduce bias, resampling methods have been applied to balance the class distribution are classified into two groups: oversampling and undersampling methods [25].”
MethodPCCT: Progressive Class-Center Triplet Loss for Imbalanced Medical Image Classification
Chen, K., Lei, W., Zhao, S., Zheng, W., & Wang, R. (2023) · IEEE Journal of Biomedical and Health Informatics · doi:10.1109/jbhi.2023.3240136 · Cited by 44
“For example, the re-sampling [4] strategy can be applied to over-sample the limited data from infrequent classes or undersample the data from frequent classes to balance training data across classes.”
BackgroundZong, Y., Nian, Y., Zhang, C., Tang, X., Wang, L., & Zhang, L. (2024) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2024.109580 · Cited by 41
A Survey of Imbalanced Learning on Graphs: Problems, Techniques, and Future Directions
Liu, Z., Yuan, L., Chen, N., Wang, Q., Hooi, B., & He, B. (2025) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2025.3549299 · Cited by 66
“Specifically, the model tends to be well-trainedon high-resource groupsthathaveampledata,while underperforming on low-resource groups that are data-limited, leading to unclear class boundaries [14], [15], [16].”
Detection of Myocardial Infarction Using ECG and Multi-Scale Feature Concatenate
Jian, J.-Z., Ger, T.-R., Lai, H.-H., Ku, C.-M., Chen, C.-A., Abu, P. A. R., & Chen, S. (2021) · Sensors · doi:10.3390/s21051906 · Cited by 41
“This was achieved through random under sampling, a data level solution for imbalanced data [51].”
BackgroundAnalyzing the Usefulness of the DARPA OpTC Dataset in Cyber Threat Detection Research
Anjum, M. M., Iqbal, S., & Hamelin, B. (2021) · ACM Symposium on Access Control Models and Technologies (SACMAT) · doi:10.1145/3450569.3463573 · Cited by 56
“Specifically, it becomes difficult to train deep learning models on the imbalanced dataset since the performance of deep learning models in terms of generalization tend to degrade with imbalanced class representation [5, 11].”
BackgroundDeep learning for white cabbage seedling prediction
Perugachi-Diaz, Y., Tomczak, J. M., & Bhulai, S. (2021) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2021.106059 · Cited by 42
Nakamura, Y., Hanaoka, S., Nomura, Y., Nakao, T., Miki, S., Watadani, T., Yoshikawa, T., Hayashi, N., & Abe, O. (2021) · BMC Medical Informatics and Decision Making · doi:10.1186/s12911-021-01623-6 · Cited by 40
“Oversampling positive data can be one of the methods to minimize the negative impact of the class imbalance [47].”
BackgroundAbdolahi, M., Salehi, M., Shokatian, I., & Reiazi, R. (2020) · Medical Journal of the Islamic Republic of Iran · doi:10.47176/mjiri.34.140 · Cited by 44
Few-Shot Class Incremental Learning Leveraging Self-Supervised Features
Ahmad, T., Dhamija, A. R., Cruz, S., Rabinowitz, R., Li, C., Jafarzadeh, M., & Boult, T. E. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw56347.2022.00435 · Cited by 41
“In [5], a comprehensive evaluation of recently proposed incremental learning approaches [2–4,8,25,26,30,36,42,48,56] was conducted.”
BackgroundVillon, S., Iovan, C., Mangeas, M., & Vigliola, L. (2022) · Sensors · doi:10.3390/s22020497 · Cited by 43
“The second approach to tackle the long-tail dataset issue consists in accounting for the imbalance of samples for each class in the training algorithms itself [45].”
An Improved Lightweight Network for Real-Time Detection of Apple Leaf Diseases in Natural Scenes
Liu, S., Qiao, Y., Li, J., Zhang, H., Zhang, M., & Wang, M. (2022) · Agronomy · doi:10.3390/agronomy12102363 · Cited by 47
“Using the “LabelImg” annotation tool [28], areas of diseased leaves in the image can be accurately In deep learning, the effect of class imbalance and small number of samples on recognition performance is detrimental [29]. marked.”
MethodDual-Branch Hybrid Learning Network for Unbiased Scene Graph Generation
Zheng, C., Gao, L., Lyu, X., Zeng, P., Saddik, A. E., & Shen, H. T. (2023) · IEEE Transactions on Circuits and Systems for Video Technology · doi:10.1109/tcsvt.2023.3297842 · Cited by 41
“It is divided into two types: under-sampling [45], [46] and over-sampling [46], [47].”
BackgroundA Survey of Tax Risk Detection Using Data Mining Techniques
Zheng, Q., Xu, Y., Liu, H., Shi, B., Wang, J., & Dong, B. (2023) · Engineering · doi:10.1016/j.eng.2023.07.014 · Cited by 48
Predicting financial distress using machine learning approaches: Evidence China
Rahman, M. J., & Zhu, H. (2024) · Journal of Contemporary Accounting & Economics · doi:10.1016/j.jcae.2024.100403 · Cited by 50
Imbalanced deep transfer network for fault diagnosis of high-speed train traction motor bearings
Liu, Y., Li, X., Zhang, X., Fan, L., Chen, X., & Gong, B. (2024) · Knowledge-Based Systems · doi:10.1016/j.knosys.2024.111682 · Cited by 40
Leveraging GANs data augmentation for imbalanced medical image classification
Ding, H., Huang, N., & Cui, X. (2024) · Applied Soft Computing · doi:10.1016/j.asoc.2024.112050 · Cited by 40
A comprehensive review of research on surface defect detection of PCBs based on machine vision
He, Z., Lian, Y., Wang, Y., & Lu, Z. (2025) · Results in Engineering · doi:10.1016/j.rineng.2025.106437 · Cited by 50
Riid, A., Lõuk, R., Pihlak, R., Tepljakov, A., & Vassiljeva, K. (2019) · Applied Sciences · doi:10.3390/app9224829 · Cited by 40
“It is known that class imbalance during training reduces the performance of deep neural networks [87].”
BackgroundChallenges and Prospects in Vision and Language Research
Kafle, K., Shrestha, R., & Kanan, C. (2019) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2019.00028 · Cited by 37
“to class imbalance and dataset bias (Godil et al., 2014; Buda et al., 2018).”
Optimal hyperparameter selection of deep learning models for COVID-19 chest X-ray classification
Adedigba, A. P., Adeshina, S. A., Aina, O. E., & Aibinu, A. M. (2021) · Intelligence-Based Medicine · doi:10.1016/j.ibmed.2021.100034 · Cited by 44
“Popular techniques for handling class imbalance include the following: random minority oversampling, random majority undersampling, and cost-sensitive learning [16].”
A simple graph-based semi-supervised learning approach for imbalanced classification
Deng, J., & Yu, J.-G. (2021) · Pattern Recognition · doi:10.1016/j.patcog.2021.108026 · Cited by 40
Curiosity-Driven Experience Prioritization via Density Estimation
Zhao, R., & Tresp, V. (2019) · arXiv · doi:10.48550/arxiv.1902.08039 · Cited by 47
“To overcome the class imbalance issue in supervised learning, such as training deep convolutional neural networks with biased datasets, researchers utilized over-sampling and under-sampling techniques [14, 7, 16].”
Malik, K., Robertson, C., Braun, D. C., & Greig, C. (2021) · International Journal of Applied Earth Observation and Geoinformation · doi:10.1016/j.jag.2021.102510 · Cited by 41
Yang, M., Wang, W., Gao, Q., Zhao, C., Li, C., Yang, X., Li, J., Li, X., Cui, J., Zhang, L., et al. (2022) · Environmental Science and Pollution Research · doi:10.1007/s11356-022-23280-6 · Cited by 40
“Since the negative impact of category imbalance on classification performance would emerge with the increase of the tasks size (Buda et al. 2018), image augmentation processing on the dataset is needed to expand the number of small phytoplankton samples to improve the accuracy of the model and…”
BackgroundA Swin Transformer-based model for mosquito species identification
Zhao, D.-Z., Wang, X.-K., Zhao, T., Li, H., Xing, D., Gao, H., Song, F., Chen, G., & Li, C.-X. (2022) · Scientific Reports · doi:10.1038/s41598-022-21017-6 · Cited by 39
GRTR: Gradient Rebalanced Traffic Sign Recognition for Autonomous Vehicles
Guo, K., Wu, Z., Wang, W., Ren, S., Zhou, X., Gadekallu, T. R., Luo, E., & Liu, C. (2023) · IEEE Transactions on Automation Science and Engineering · doi:10.1109/tase.2023.3270202 · Cited by 41
“Recently, many workers have started to apply deep learning techniques to long-tailed datasets and have proposed many solutions, which can be divided into data balancing methods [45], [46] and network tuning methods [47].”
BackgroundDeep learning in material recovery: Development of method to create training database
Vrancken, C., Longhurst, P., & Wagland, S. T. (2019) · Expert Systems with Applications · doi:10.1016/j.eswa.2019.01.077 · Cited by 44
“The mean classification accuracy of 77.5% reached levels of performance typically obtained with databases containing a similar number of training images (Bell et al., 2015; Buda et al., 2017).”
BackgroundInfluentialGao, X., & Wang, X. (2019) · International Journal of Computer Assisted Radiology and Surgery · doi:10.1007/s11548-019-02070-5 · Cited by 39
The PAU Survey: Photometric redshifts using transfer learning from simulations
Eriksen, M., Alarcon, A., Cabayol, L., Carretero, J., Casas, R., Castander, F. J., De Vicente, J., Fernandez, E., Garcia-Bellido, J., Gaztanaga, E., et al. (2020) · Monthly Notices of the Royal Astronomical Society · doi:10.1093/mnras/staa2265 · Cited by 39
VistrongerDet: Stronger Visual Information for Object Detection in VisDrone Images
Wan, J., Zhang, B., Zhao, Y., Du, Y., & Tong, Z. (2021) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw54120.2021.00316 · Cited by 42
“balanced the differences between long-tail categories and head categories with more samples for tail classes during training [2].”
BackgroundMethodInfluentialUsing deep convolutional neural networks to forecast spatial patterns of Amazonian deforestation
Ball, J., Petrova, K., Coomes, D. A., & Flaxman, S. R. (2022) · Methods in Ecology and Evolution · doi:10.1111/2041-210x.13953 · Cited by 50
Class-Imbalanced Semi-Supervised LearningPreprint
Hyun, M., Jeong, J., & Kwak, N. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2002.06815 · Cited by 63
“Data-level methods approach the problem by over-sampling minor classes or under-sampling major classes (Masko & Hens-man, 2015; Lee et al., 2016; Pouyanfar et al., 2018; Buda et al., 2018).”
BackgroundMethodGao, H., Zhang, Y., Chen, Z., Xu, F., Hong, D., & Zhang, B. (2023) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2023.3288739 · Cited by 38
“Previous studies indicated that class imbalance has a significant detrimental effect on model convergence and generalization of the deep network [49], [50].”
BackgroundDynamic Residual Classifier for Class Incremental Learning
Chen, X., & Chang, X. (2023) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.01718 · Cited by 42
“The adjusted losses [36, 8, 46] and the data re-samling [1, 2, 12, 28] are two kinds of important techniques for Long-tailed Recognition.”
MethodThe Intelligent Design of Silicon Photonic Devices
Li, Z., Zhou, Z., Qiu, C., Chen, Y., Liang, B., Wang, Y., Liang, L., Lei, Y., Song, Y. X., Jia, P., et al. (2024) · Advanced Optical Materials · doi:10.1002/adom.202301337 · Cited by 38
Stadelmann, T., Amirian, M., Arabaci, I., Arnold, M., Duivesteijn, G. F., Elezi, I., Geiger, M., Lörwald, S., Meier, B. B., Rombach, K., et al. (2018) · Lecture notes in computer science · doi:10.1007/978-3-319-99978-4_2 · Cited by 40
“Using pre-trained networks and applying transfer learning [41] – Increasing the weight of the minority class in the optimization loss [8] – Generating synthetic data for the minority class using SMOTE [11] or GANs [21] Selecting a suitable data augmentation approach according for the task is a…”
MethodDing, H., Xia, W., Zhang, L., Mao, Q., Cao, B., Zhao, Y., Xu, L., Jiang, F., & Dong, G. (2020) · Frontiers in Oncology · doi:10.3389/fonc.2020.01186 · Cited by 38
“However, an imbalance in the amount of data will adversely affect the performance of a deep learning model (18).”
Background
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
