Long sequence temporal knowledge tracing for student performance prediction via integrating LSTM and informer

Area of Science:

• Educational Data Mining
• Machine Learning in Education
• Artificial Intelligence in Education

Background:

• Knowledge tracing (KT) models student learning performance using historical data.
• Existing KT models often struggle with long sequences of student interactions.
• There is a need for advanced methods capable of handling long-term dependencies in educational data.

Purpose of the Study:

• To develop a robust pipeline for long-sequence time-series forecasting in knowledge tracing.
• To improve the accuracy and efficiency of predicting student knowledge states over extended periods.
• To address the limitations of current KT approaches in handling long-term educational data.

Main Methods:

• A bidirectional LSTM model was used for embedding exercise-answering records.
• Time stamps and student exercise data were combined into input vectors.
• An Informer model with a probability-sparse self-attention mechanism processed sequential data.
• Temporal information and individual knowledge states were integrated for exercise prediction.

Main Results:

• The proposed LSTKT model demonstrated significant quantitative improvements over state-of-the-art KT algorithms.
• On the Assistments2009 dataset, the model achieved 78.49% accuracy and 78.81% AUC.
• On the Assistments2017 dataset, accuracy reached 74.22% and AUC reached 72.82%.
• On the EdNet dataset, the model attained 68.17% accuracy and 70.78% AUC.

Conclusions:

• The developed long-sequence time-series forecasting pipeline effectively enhances knowledge tracing capabilities.
• The Informer model's probability-sparse self-attention mechanism efficiently handles long sequences.
• The LSTKT model offers a promising advancement for educational data mining and personalized learning.