In 5-fold cross-validation, multi-bin time-variant model fits outperform single-bin fits, with diminishing returns beyond a 5-bin split and further gains from differentiating by learner and amount of practice¶

Claim · All claims
Evidence · 1 study (2 entries) · associational r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Dividing the interval range into multiple time bins generally produced better out-of-sample predictive performance than single-bin variants, with diminishing returns beyond a 5-bin split. → Maarten van der Velde 2026 q2 i? Differentiating fits by learner and by number of practice trials in the first session further improved predictive performance. → Maarten van der Velde 2026 (2)

Evidence¶

Maarten van der Velde 2026¶

Maarten van der Velde, Florian Sense, Jelmer P. Borst & Hedderik van Rijn. (2026). Explaining forgetting at different timescales requires a time-variant forgetting function. npj Science of Learning. https://doi.org/10.1038/s41539-026-00447-1

q2 · i? · associational · r2

5-fold cross-validation with stratification by learner and interval bin, comparing summed log-likelihood on held-out folds; 3,650 sequences (14.1%) were excluded, leaving 22,193 sequences for evaluation. "a division into multiple time bins generally resulted in a better-performing model than the single-bin variants."

"As Fig.2 shows, a division into multiple time bins generally resulted in a better-performing model than the single-bin variants, albeit with diminishing returns beyond a 5-bin split."

Maarten van der Velde 2026 (2)¶

Maarten van der Velde, Florian Sense, Jelmer P. Borst & Hedderik van Rijn. (2026). Explaining forgetting at different timescales requires a time-variant forgetting function. npj Science of Learning. https://doi.org/10.1038/s41539-026-00447-1

q2 · i? · associational · r2

Same cross-validated comparison at the group, by-learner, and by-practice-trial level. The article reports that "differentiating between learners, as well as differentiating between different amounts of practice prior to the between-session interval, produces a further improvement in predictive performance."

"As Fig. 2a shows, differentiating between learners, as well as differentiating between different amounts of practice prior to the between-session interval, produces a further improvement in predictive performance."

Discussion¶

Related Claims¶

• Allowing ACT-R memory parameters to vary with the between-session interval produces a strong fit between model and recall data across timescales — related
• Extrapolating from parameters fitted to a restricted range of between-session intervals produces worse predictions than time-variant fits — related
• Optimal ACT-R parameter values change systematically and predictably with the length of the between-session interval — related