Improving Imbalanced Classification Based on Convolutional Neural Network Model with Reinforcement Learning

Imbalanced data sets across different group can significantly affect the performance of standard algorithms in classification tasks. When such data sets are used without preprocessing to ensure that each class has a similar size, the accuracy of the model can become biased. Specifically, the model will perform poorly on underrepresented classes while being accurate on classes with more data. This study employs a neural network model with reinforcement learning to address the issue of data imbalance. The mod el's performance is tested using the MNIST image dataset, which contains 70,000 images of digits from 0 to 9. In this research, balanced data is simulated to be imbalanced by randomly reducing the number of instances in class 5 compared to the other classes. The experiments are divided into two sets: the first set has a data imbalance ratio (IR) ranging from 1 to 15, increasing by 1 at each step, and the second set has an IR ranging from 1 to 40, increasing by 5 at each step. The results indicate that the proposed model effectively addresses high imbalance ratios in data classification compared to standard models.