File size: 1,521 Bytes
29baa44 ae3d87c 29baa44 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | """
Forgetting model used for spaced repetition.
Recall probability decays exponentially, R = exp(-days / S), where the memory
stability S (in days) grows with successful practice and mastery, the same shape as
half-life regression (Settles & Meeder, 2016) with hand-set coefficients.
Forgetting is applied at *read* time to produce an "effective" mastery; it is never
written back to the database, so repeated reads don't compound decay.
"""
import math
from datetime import datetime
REVIEW_RECALL_THRESHOLD = 0.7
def stability_days(practice_count, correct_count, mastery):
correct_count = correct_count or 0
practice_count = practice_count or 0
wrong = max(0, practice_count - correct_count)
return max(0.5, 1.0 + 1.6 * correct_count ** 0.85 - 0.3 * wrong ** 0.7) * (0.5 + mastery)
def recall_probability(last_practiced, practice_count, correct_count, mastery, now=None):
if not last_practiced or not practice_count:
return 1.0
now = now or datetime.utcnow()
days = max(0.0, (now - last_practiced).total_seconds() / 86400)
return math.exp(-days / stability_days(practice_count, correct_count, mastery))
def effective_mastery(mastery, recall):
# Forgotten knowledge is easier to relearn than new material, so it never decays to zero.
return mastery * (0.35 + 0.65 * recall)
def days_until_review(practice_count, correct_count, mastery, threshold=REVIEW_RECALL_THRESHOLD):
return -math.log(threshold) * stability_days(practice_count, correct_count, mastery)
|