Reducing Overfitting Risk in Small-Sample Learning with ANN: A Case of Predicting Graduate Admission Probability

AI-assisted personal educational career planning holds immense promise, especially with artificial neural network algorithms demonstrating significant potential in predicting graduate admission probabilities for specific schools. However, issues like graduate admission probability prediction are typical instances of small-sample learning problems. Overcoming neural network overfitting in small-sample learning is a major challenge. To address this, we first introduce a method that cyclically runs shallow neural networks to avoid the traditional use of a high number of epochs called Cyclic-LightNet(CLN). Secondly, in previous research, a validation set is typically employed to promptly detect overfitting. However, this approach further depletes the actual training data in the dataset. Consequently, we employ a dynamic training method during the training process called Dynamic Training Set Augmentation(DTSA), supplementing the training set with the validation set. This method effectively mitigates the risk of overfitting. Finally, we put CLN and DTSA into practice on publicly available small-sample datasets, achieving precise predictions of graduate admission probabilities. Experimental results demonstrate a significant improvement in prediction accuracy and effective resistance against overfitting risk.