The problem
Digital materials record actions such as moving to the next page, returning to a previous page and opening a resource. The study asks how these logs can inform the selection of actions to recommend to a learner.
The research idea
Counts of NEXT, PREV and OPEN actions feed a Transformer that predicts five grade categories. Two methods select actions to modify: one uses loss gradients, and the other uses attention differences relative to higher-performing learners.
What was evaluated
The experiments modify action counts in input logs and submit them to the same trained grade predictor. They compare changes in predicted grades against randomly selected actions.
Considering applications
The methods offer a way to generate recommendation candidates for learning-support systems. Further evaluation can connect the educational meaning of recommended actions with observed learning outcomes.
Original research
Recommending Learning Actions Using Neural Network
Kohama, Hirokazu, Ban, Yuki, Hirakawa, Tsubasa, Yamashita, Takayoshi, Fujiyoshi, Hironobu, Itai, Akitoshi, Usami, Hiroyasu
Proceedings of the 31st International Conference on Computers in Education, 2023 · ICCE 2023