5D Library › 诊学一体转化漏斗、掌握度门控课程与持牌人网络经济学 · English
第四章 章节粒度、流失风险与十二道门槛
章节切多细
在线课程的完课率一向很低。课程设计者能控制的变量之一是章节粒度:同样 204 学时,可以切成 12 个大章,也可以切成 400 个小节。本章用一个两项风险模型来回答切多细。第一项是断点风险:每个章节结束都是一扇「门」,学员有固定的概率 从这里离开。第二项是疲劳风险:章节越长,学员在章节中途放弃的风险越大,设为 ( 为一章的学时)。切成 章时,学完的概率为
在 处取得最大值 。在章末加一场考试(门槛)还要再加一项风险 。
注
风险参数是示意值。 、、 只用来展示模型的形状;真实的参数需要用课程的逐章流失数据估计。
推导
# Chapter 4. [p07] Chapter granularity versus dropout hazard.
# H = 204 hours of content (p03) are cut into n chapters of h = H / n
# hours. Two hazards act on a learner: a fixed chance a of quitting at
# every chapter boundary (each break is a door out), and a fatigue
# hazard that grows with chapter length, b h^2 per chapter. Survival to
# the end is
# S(n) = exp(-n a - n b h^2) = exp(-n a - b H^2 / n),
# maximised at n* = H sqrt(b / a), where S* = exp(-2 H sqrt(a b)).
# A gate (an exam) at a boundary adds hazard c; gating every chapter
# multiplies it by n, gating once per subject by 12.
import math
H, A, B, C = 204.0, 0.004, 0.0006, 0.010
def surv(n, gates):
h = H / n
return math.exp(-n * A - n * B * h * h - gates * C)
if __name__ == "__main__":
n_star = H * math.sqrt(B / A)
print(" n* = {:.1f} chapters of {:.2f} h; S* = {:.4f}".format(
n_star, H / n_star, math.exp(-2 * H * math.sqrt(A * B))))
print(" n hours/ch S no gates S gate/subject S gate/chapter")
for n in (12, 24, 48, 79, 120, 204, 408):
print(" {:5d} {:10.2f} {:12.4f} {:15.4f} {:15.4f}".format(
n, H / n, surv(n, 0), surv(n, 12), surv(n, n)))
best = max(range(1, 600), key=lambda n: surv(n, n))
print(" with a gate on every chapter the best n is {} ({:.4f})".format(
best, surv(best, best)))
print("\nReading: the optimum is flat - anything from about 48 to")
print("120 chapters keeps survival within a few points of S*. Very")
print("long chapters cost most; gating every short chapter costs")
print("far more than gating once per subject.")
# n* = 79.0 chapters of 2.58 h; S* = 0.5315
# n hours/ch S no gates S gate/subject S gate/chapter
# 12 17.00 0.1190 0.1055 0.1055
# 24 8.50 0.3210 0.2847 0.2525
# 48 4.25 0.4906 0.4351 0.3036
# 79 2.58 0.5315 0.4714 0.2412
# 120 1.70 0.5025 0.4457 0.1514
# 204 1.00 0.3913 0.3470 0.0509
# 408 0.50 0.1839 0.1631 0.0031
# with a gate on every chapter the best n is 42 (0.3065)
#
# Reading: the optimum is flat - anything from about 48 to
# 120 chapters keeps survival within a few points of S*. Very
# long chapters cost most; gating every short chapter costs
# far more than gating once per subject.最优是切成约 79 章,每章 2.58 小时,完课率 0.5315。最优点很平:48 到 120 章之间,完课率都不低于 0.4906(120 章时为 0.5025)。真正伤人的是两端:切成 12 个大章(每章 17 小时)时完课率只有 0.1190;切成 408 个半小时的小节时只剩 0.1839。
门槛的位置更要紧。每个科目末考一次(12 次),79 章时完课率从 0.5315 降到 0.4714;若每章末都考,降到 0.2412。章章都考时,最优的切法变成 42 章,完课率 0.3065。换言之,「掌握度门控」应当按科目设门,而不是按章节设门:短章节用来降低疲劳,门槛用来确认掌握,两者不要叠在一起。
一个队列走完 12 道门槛
现在把第二、三章的部件装在一起:10,000 位学员按第三章的人群分布 Beta(6, 3) 取基础掌握度,按第二章的拓扑顺序学 12 个科目,每章 2.58 小时(第七段程序的流失风险)。学完一科后掌握度提高 0.08 左右。在门控方案下,每科结束考第三章的 5 道题,不及格再加 0.15 的离开概率,最多补考 3 次,每次补考掌握度再提高 0.03。在开放方案下没有考试,看完就算完成。
推导
# Chapter 4. [p08] A cohort through 12 gated subjects.
# 10,000 learners with base mastery theta ~ Beta(6, 3) study the 12
# subjects of p03 in topological order, in chapters of 2.58 hours
# (p07: hazard 0.004 per boundary plus 0.0006 h^2 per chapter). Study
# lifts mastery on a subject to theta + 0.08 + noise (sd 0.06, capped
# at 0.97). Under the GATE regime each subject ends with the exam of
# p05 (5 x 4 points, pass 16, k = 6); a failure costs a further 0.15
# chance of quitting, and up to 3 retakes add 0.03 mastery each. Under
# the OPEN regime there is no exam: a learner finishes by watching.
# Reported: completion, mean base mastery of finishers, and the share
# of finishers with base mastery >= 0.80.
import math
import random
ORDER = ["astro", "lifa", "bazi", "chenggu", "liuyao", "meihua",
"numer", "qimen", "fengshui", "uranian", "xingming", "ziwei"]
HOURS = {"lifa": 12, "numer": 6, "astro": 14, "bazi": 30, "liuyao": 18,
"ziwei": 28, "chenggu": 4, "meihua": 10, "qimen": 32,
"fengshui": 24, "xingming": 10, "uranian": 16}
A, B, CH, K = 0.004, 0.0006, 2.58, 6.0
def exam(rng, m):
s = 0
for _ in range(5):
q = rng.betavariate(K * m, K * (1 - m))
s += sum(rng.random() < q for _ in range(4))
return s >= 16
def run(rng, gate, n=10000):
done, base = 0, []
tries = 0
for _ in range(n):
th = rng.betavariate(6, 3)
alive = True
for subj in ORDER:
nch = HOURS[subj] / CH
if rng.random() > math.exp(-nch * (A + B * CH * CH)):
alive = False
break
if not gate:
continue
m = min(0.97, max(0.05, th + 0.08 + rng.gauss(0, 0.06)))
for t in range(4):
tries += 1
if exam(rng, m):
break
if t == 3 or rng.random() < 0.15:
alive = False
break
m = min(0.97, m + 0.03)
if not alive:
break
if alive:
done += 1
base.append(th)
return done / n, sum(base) / len(base), sum(
t >= 0.80 for t in base) / len(base), tries / n
if __name__ == "__main__":
rng = random.Random(74008)
print(" regime completion mean theta share >= 0.80 exams/learner")
for name, gate in (("OPEN", False), ("GATE", True)):
c, mt, hi, ex = run(rng, gate)
print(" {:6s} {:11.4f} {:11.4f} {:14.4f} {:14.2f}".format(
name, c, mt, hi, ex))
print("\nReading: twelve gates turn a course that half the cohort")
print("finishes into one that about a sixth finishes; those who do")
print("are much stronger. The gate is a filter first and a teacher")
print("second, and its price is paid in completion.")
# regime completion mean theta share >= 0.80 exams/learner
# OPEN 0.5307 0.6681 0.2093 0.00
# GATE 0.1656 0.8126 0.5694 6.89
#
# Reading: twelve gates turn a course that half the cohort
# finishes into one that about a sixth finishes; those who do
# are much stronger. The gate is a filter first and a teacher
# second, and its price is paid in completion.开放方案下完课率 0.5307,完课者的平均基础掌握度 0.6681,与人群几乎相同,其中只有 0.2093 不低于 0.80。门控方案下完课率降到 0.1656,但完课者的平均基础掌握度是 0.8126,其中 0.5694 不低于 0.80;平均每位学员参加了 6.89 次考试。
这说明门控首先是一个过滤器,其次才是一个教学手段。在本模型中,门控带来的「提高」主要是把弱学员筛掉,而不是把他们教强:每次补考只加 0.03,远不足以把一位弱学员提高到胜任水平。掌握学习的文献报告过可观的效果 (Kulik et al. 1990),Bloom 甚至提出一对一辅导可以提高两个标准差 (Bloom 1984);但这些效果来自门槛之后的矫正教学,不是门槛本身。一个只设门而不提供矫正的课程,得到的是一批强学员和大量流失。
假说 · 尚未验证
作者的诠释(尚未验证): 在作者的课程中,若每次未通过之后都提供针对错题的矫正单元(而不只是重考),门控方案的完课率可以显著高于 0.1656,同时完课者的掌握度不低于 0.80。什么会证明这一判断错:若在预注册实验中,「门控 + 矫正」与「门控 + 直接重考」的完课率与第 90 天迁移测验成绩都没有差别,那么矫正单元不起作用,门控只是过滤器。
小结
章节粒度有一个平坦的最优,约 79 章、每章 2.58 小时;太长与太短都伤完课率。门槛应按科目而不是按章节设置。12 道科目门槛把完课率从 0.5307 降到 0.1656,完课者强得多;门控在本模型中首先是过滤,其次才是教学。
参考文献
Fan, Y. (2026). Diagnostic-to-Pedagogy Conversion Funnels, Mastery-Gated Curriculum Architecture and Certification Network Economics: Self-Cannibalization Bounds, Spaced-Retrieval Curves and Two-Sided Practitioner Licensing [诊学一体转化漏斗、掌握度门控课程与持牌人网络经济学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090197 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).