Use early student-model predictions to identify at-risk students and adapt pedagogical strategy during tutoring¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The article's findings yield a learning environment that can "foretell students' performance and learning gains early, and can render adaptive pedagogical strategy accordingly". Because BKT+SK predicts post-test scores from the earliest 50% of interaction sequences and LSTM predicts learning gains from the earliest 70%, an ITS can identify students at high risk of failing the post-test or of low learning gain well before training ends and offer adaptive recommendation at that point.

Design Implications¶

Context¶

Requirements¶

• A trained student model (BKT+SK for post-test, LSTM for learning gains) and logged student-ITS interaction trajectories including skills, interventions, performance and response speed

Constraints¶

• Validated on two datasets involving both elicit and tell interventions; the authors state it is not clear whether the same conclusions hold for datasets with only one intervention type or other intervention types such as skip and justify

Target Learners¶

• college students in intelligent tutoring systems for physics or probability

Target Learning Goals¶

• early identification of students at risk of low post-test scores or low learning gains

Related Strategies¶

• Use predicted hint-taking likelihood and hint effects to adaptively decide whether to withhold or provide hints

Examples¶

-

Key Sources¶

• Ye Mao, Chen Lin, and Min Chi. (2018). Deep Learning vs. Bayesian Knowledge Tracing: Student Models for Interventions. Journal of Educational Data Mining, Volume 10, No 2. https://jedm.educationaldatamining.org/index.php/JEDM/article/view/318