Improving the Performance of the Convolutional Neural Network Using Incremental Weight loss Function to Deal with Class Imbalanced Data

Class-imbalanced datasets are common in many real-world domains, such as health, banking, and security. Machine learning researchers have recently focused on the classification of such datasets, where the costs of different types of misclassifications are unequal, the classes have different prior probabilities, or both. The performance of most standard classifier learning algorithms is significantly affected by class imbalance, where the algorithms are often biased toward the majority class instances despite recent advances in deep learning. However, there is very little empirical work on deep learning with class imbalance.To address this issue, we propose an incremental weighted cross entropy loss function. The proposed method involves gradually increasing the weight of the minority class as the training progresses, until it reaches the specified amount at the end of the training. Through experiments, we demonstrate the convergence and efficiency of the proposed method. The results of experiments on three datasets, including artificial datasets, human activity recognition dataset, and CIFAR-10, demonstrate the convergence and performance of the proposed method. The proposed method is compared with decision tree-based AdaBoost, Cross Entropy-based convolutional neural network, weighted Cross Entropy -based CNN, SMOTE method, and ensemble CNNs method. With accuracy gains of 94.6%, 92.92%, and 69.23% on the three datasets (CIFAR-10 with 5% imbalance rate), the proposed method outperformed the other methods. Additionally, the accuracy on the artificial dataset was 17.77% higher than the traditional decision tree-based AdaBoost method.