Authors:
Jakub Pokrywka
1
;
Marcin Biedalak
2
;
Filip Graliński
1
and
Krzysztof Biedalak
2
Affiliations:
1
Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Poznań, Poland
;
2
SuperMemo World, Poland
Keyword(s):
Spaced Repetition, LSTM, Metalearning.
Abstract:
Spaced repetition is a human learning technique focused on optimizing time intervals between a student’s repetitions of the same information items. It is designed for the most effective long-term high-retention knowledge acquisition in terms of a student’s time spent on learning. Repetition of an information item is performed when its estimated recall probability falls to the required level. Spaced repetition works particularly well for itemized knowledge in areas requiring high-volume learning like languages, computer science, medicine, etc. In this work, we present a novel machine-learning approach for the prediction of recall probability developed using the massive repetition data collected in the SuperMemo.com learning ecosystem. The method predicts the probability of remembering an item by a student using an LSTM neural network. In our experiments, we observed that applying the spaced repetition research expert algorithms (Woźniak et al., 2005), like imposing the negative expone
ntial function as the output forgetting curve, increases the LSTM model performance. We analyze how this model compares to other machine-learning or expert methods such as the Leitner method, XGBoost, half-life regression, and the spaced repetition expert algorithms. We found out that the choice of evaluation metric is crucial. Furthermore, we elaborate on this topic, finally selecting macro-average MAE and macro-average Likelihood for the primary and secondary evaluation metrics.
(More)