Skip to main content

Table 17 Summary of deep learning class imbalance methods

From: Survey on deep learning with class imbalance

Method Network type Method type Description
ROS [23, 79] CNN Data ROS of minority classes until class balance is achieved
RUS [23] CNN Data RUS of majority classes until class balance is achieved
Two-phase learning [20, 23] CNN Data Pre-training with RUS or ROS, then fine-tuning with all data
Dynamic sampling [21] CNN Data Sampling rates adjust throughout training based on previous iteration’s class-wise F1-scores
MFE and MSFE loss [18] MLP Algorithm New loss functions allow positive and negative classes to contribute to loss equally
Focal loss [88, 103] CNN Algorithm New loss function down-weights easy-to-classify samples, reducing their impact on total loss
CSDNN [89] MLP Algorithm CE loss function modified to incorporate a pre-defined cost matrix
CoSen CNN [19] CNN Algorithm Cost matrix is learned through backpropagation and incorporated into output layer
CSDBN-DE [90] DBN Algorithm Cost matrix is learned through evolutionary algorithm and incorporated into output layer
Threshold moving [23] CNN Algorithm Decision threshold is adjusted by dividing output probabilities by prior class probabilities
Category centers [91] CNN Algorithm Class centroids are calculated in deep feature space and K-NN method discriminates
Very-deep NNs [92] CNN Algorithm CNN network depths of up to 50 layers are used to examine convergence rates
LMLE [22] CNN Hybrid Triple-header hinge loss and quintuplet sampling generate more discriminative features
DOS [117] CNN Hybrid Minority class over-sampled in deep feature space using K-NN and micro-cluster loss
CRL loss [118] CNN Hybrid Class Rectification loss and hard sample mining produce more discriminative features