In Mnemosyne log data, item-specific difficulty parameters outperform a global difficulty for lower and higher Leitner decks, while global difficulty performs better for intermediate decks.¶

Claim · All claims
Evidence · 1 study · causal r? · q2
Currency · newest study 2016 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person · overdue for a recheck since 2019: tool and model accuracy changes quickly

Subclaims¶

q2 i? Item-specific difficulties outperformed a global item difficulty for decks q ≤ 2 and q > 5, but the global difficulty performed better for intermediate decks; no numeric AUC values are printed. → Reddy 2016

Evidence¶

Reddy 2016¶

Reddy, S., Labutov, I., Banerjee, S., & Joachims, T. (2016). Unbounded Human Learning: Optimal Scheduling for Spaced Repetition. KDD ’16, San Francisco, CA, USA. https://doi.org/10.1145/2939672.2939850

q2 · i? · causal · r?

Observational study of Mnemosyne flashcard log data, comparing memory models by cross-validated validation AUC per deck bin. "Item-specific difficulties θi outperform global item difficulty θ for lower decks" and higher decks, but the global parameter did better for intermediate decks (models 5 vs. 10, 8 vs. 13).

"Item-specific difficulties θi outperform global item difficulty θ for lower decks (qij≤ 2) and higher decks ( qij > 5), but the global dif- ficulty performs better for intermediate decks (model 5 vs. 10, 8 vs. 13)"

Discussion¶

Related Claims¶

• Memory-decay-based models fit fact-learning datasets better than models insensitive to memory decay — the general claim this finding bears on, with the studies behind it

• Under the Leitner Queue Network model, raising the rate of new-item introduction beyond a threshold causes a phase transition in learning rate, which a Mechanical Turk vocabulary experiment reproduced. — the general claim this finding bears on, with the studies behind it

• In Mnemosyne flashcard log data, adding a delay term improves the recall-prediction performance of exponential forgetting curve memory models. — related
• In Mnemosyne log data, setting memory strength equal to an item's Leitner deck position predicts recall better than number of past reviews, which beats constant strength. — related
• In Mnemosyne log data, exponential forgetting curve models that include a delay term perform comparably to 1PL-IRT, the best-performing benchmark model. — related
• Correlation between log response time and response likelihood is negative for almost all items, and its magnitude depends on the item difficulty parameter — related
• A one-parameter IRT model outperforms global and per-pitch baselines on AUC, BCE and Brier score but not on threshold accuracy for held-out singing attempts — related
• Item difficulty was best predicted using signal from the item stem — related
• All models predict performance worst for students of intermediate overall ability, with best performance for high and low performing students — related