Colearn: a multi-task memory-augmented deep learning model jointly predicting knowledge state and hint-taking¶

Theory · All theories
Evidence · no claims cited

Description¶

Colearn extends the DKVMN memory-augmented knowledge tracing model by encoding each interaction as a tuple (qt, rt, ht) of question, response, and hint usage, and by adding hint-taking prediction as an auxiliary task sharing network weights. The input encoding uses a vector of length 2|Q|+1 whose "last dimension of the vector is a binary value indicating whether a hint is taken". The loss is a weighted sum of cross-entropy losses for the two tasks with equal weights, and the network weights except the final output layer are shared between tasks.

Design Implications¶

Context¶

Requirements¶

• Requires interaction logs recording whether a hint was taken directly instead of attempting the question first, plus question identifiers and response correctness.

Constraints¶

• On Junyi, hint labels are noisy because the data records only whether a hint was one of the actions, not whether it was the first action, which the authors say reduces the benefit of incorporating hint information.

Target Learners¶

• K-12 mathematics students using interactive e-learning platforms

Target Learning Objectives¶

• Predicting future response correctness (knowledge tracing)
• Predicting likelihood of taking a hint

Claims¶

Related Theories¶

• Deep Learning Knowledge Tracing Models

Examples¶

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

Key Sources¶

• Ritwick Chaudhry, Harvineet Singh, Pradeep Dogga, and Shiv Kumar Saini. (2017). Modeling Hint-Taking Behavior and Knowledge State of Students with Multi-Task Learning. Proceedings of the 11th International Conference on Educational Data Mining. https://educationaldatamining.org/