Abstract:

Educational evaluation reform is a key component of deepening educational reform in the new era, yet traditional value-added assessment methods have technical limitations in handling dynamic characteristics and complex dependencies in the learning process. This study constructs a Temporal Pattern Attention Long Short-Term Memory neural network (TPA-LSTM) value-added assessment model based on data from 4,869 high school students in Zhejiang Province's class of 2023. By integrating quantile regression methods, the model achieves precise evaluation of temporal patterns and nonlinear changes in student performance. Through systematic analysis of Chinese language test scores across five semesters, the study examines learning trajectory characteristics at the individual level and value-added performance at the group level. The findings show that: the TPA-LSTM model achieves a root mean square error (RMSE) of 0.082 and mean absolute error (MAE) of 0.067 on the test set, significantly outperforming traditional SGP models; for students with identical scores (0.716) in the second semester of grade 11, the model identifies value-added level differences ranging from 34 to 80 based on their historical learning trajectories; and the temporal weight distribution reveals that the third and fourth semesters are critical learning periods, providing stronger interpretability for evaluation results. The study demonstrates that this model enables precise characterization of learning trajectories at the individual level while revealing developmental patterns of different student types at the group level, providing a new technical approach to improving both the predictive accuracy and educational diagnostic value of value-added assessment in educational testing.

Key words: educational testing, value-added assessment, neural network model, temporal pattern, long short-term memory network