Time series prediction of college student satisfaction based on BiLSTM in big data scenarios

Abstract

In big data scenarios, traditional LSTM (Long Short-Term Memory) only focuses on one-way sequences and cannot use future information, resulting in weak understanding of contextual information and limited ability to handle non-stationarity. When faced with massive amounts of data, the unidirectional structure of the traditional LSTM can lose some of the potential features of the time series information, making it difficult to fully capture the long-term relationships in the data. The BiLSTM (Bidirectional Long Short-Term Memory) model introduced in this paper has a bidirectional structure and stronger context understanding ability, can better identify nonlinear and non-stationary patterns, and can more effectively capture long-term dependencies. The BiLSTM model is used to predict the time series of college student satisfaction. A sliding window is designed to generate time series training samples, and feature selection and standardization are combined to improve data quality. By setting multiple hyperparameters, the model is configured with a learning rate of 0.001, a hidden layer size of 100, and a sliding window size of 10, achieving training and stable prediction effects for colleges and universities. The experimental results show that using the optimally configured BiLSTM model, the loss of the training set is reduced to 0.245, and the loss of the validation set is reduced to 0.312. The MSE is 0.1975, the MAE is 0.290, the R² is 0.89, and the time series correlation coefficient is 0.94, indicating that the model has an advantage in dealing with nonlinear, non-stationary data and long-term dependency problems. The BiLSTM model optimizes model configuration through multi-hyperparameter tuning and cross-validation, achieving efficient training and stable prediction results, providing university administrators with more accurate student satisfaction trend analysis and decision support.