A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling
Author: Junyao Ye, Jingyong Su, Yilong Cao · Type: academic · Status: draft · URL: https://dl.acm.org/doi/10.1145/3534678.3539081
Licence: ACM copyright (KDD 2022, pp. 4381-4390); paywalled full text; abstract quotable; replication data openly available on Harvard Dataverse
Summary
The peer-reviewed KDD paper from Jarrett (Junyao) Ye, the creator of FSRS, and colleagues. Using 220 million memory behavior logs from the MaiMemo language app, they built a memory model with Markov property (the DSR difficulty-stability-retrievability framework underlying FSRS) and a scheduler (SSP-MMC) that provably minimizes review cost, showing a 12.6% improvement over prior state-of-the-art schedulers. It was deployed to millions of MaiMemo learners. This is the strongest citable evidence that the modeling lineage behind FSRS is published, large-scale, and industrially validated, directly supporting Immersive’s choice of FSRS over SM-2.
Insights
- The memory-modeling research behind FSRS is peer reviewed at ACM KDD and trained on 220 million real student memory logs from a language-learning app. (medium, draft)
fsrsspaced repetitionmemoryWe have collected 220 million students’ memory behavior logs with time-series features and built a memory model with Markov property. — Abstract
- Optimized scheduling reduced review cost by 12.6% over previous state-of-the-art spaced repetition methods, meaning learners reach the same retention with measurably less review time. (medium, draft)
fsrsspaced repetitionproduct designExperimental results have shown a 12.6% performance improvement over the state-of-the-art methods. The scheduler has been successfully deployed in the online language-learning app MaiMemo to help millions of students. — Abstract