Student Academic Early Warning Prediction based on LSTM networks

In order to solve the problems of existing academic early warning systems in universities that require complete student behavior data, low prediction accuracy, and poor scalability, the author proposes an early warning model based on LSTM optimized neural network. This model mainly consists of three parts: Data processing, feature extraction, and optimization training. Data processing involves mapping student information data from low to high dimensions using RBF kernel functions, in order to preserve complete student data information and reduce the interference of data loss on the early warning system; Feature extraction refers to the classification of student information data through multidimensional normal distribution feedforward feature extraction, followed by normalization processing, and learning operations based on different categories to improve prediction accuracy; Optimization training mainly involves inputting data into LSTM neural networks optimized through adaptive excitation functions for training, obtaining graduation probabilities and weak features through probability fitting, and outputting academic prediction reports. The first year compulsory course scores, consumption information, and book borrowing frequency of 15211 graduates from a certain university were collected as data training sets, and 5301 graduates were used as test data for the experiment. The experimental results indicate that: The prediction accuracy of the model is stable at 94.1%, with a maximum of 98.06%. The average accuracy is improved by 2 percentage points compared to existing warning models, especially with a significant improvement in negative recall rate, data dependence, and result stability.