Which Features Decide? چالش 6 از جلسهی 51، سخت
تابع بازگشتی features_used(node, names) را بنویس که مجموعهی (set) اسم همهی ویژگیهایی را که در درخت پرسیده میشوند برگرداند. برگ هیچ ویژگیای ندارد. خروجی: depth 2 uses: ['sleep_hours', 'study_hours']، depth 5 uses: ['seat_row', 'sleep_hours', 'study_hours'] و check this: seat_row decides! is that fair?.
import random
names = ["study_hours", "sleep_hours", "seat_row"]
def make_students(n):
rows, labels = [], []
for i in range(n):
study = random.randint(0, 10)
sleep = random.randint(4, 10)
seat = random.randint(1, 6)
passed = study + sleep // 2 >= 8
if random.random() < 0.15: # life is noisy: 15% surprises
passed = not passed
rows.append([study, sleep, seat])
labels.append("pass" if passed else "fail")
return rows, labels
random.seed(4)
rows, labels = make_students(40) # train
trows, tlabels = make_students(40) # test
def gini(group):
if len(group) == 0:
return 0.0
score = 1.0
for c in sorted(set(group)):
p = group.count(c) / len(group)
score -= p * p
return score
def split(rows, labels, f, t):
left = [labels[i] for i in range(len(rows)) if rows[i][f] < t]
right = [labels[i] for i in range(len(rows)) if rows[i][f] >= t]
return left, right
def weighted_gini(left, right):
n = len(left) + len(right)
return len(left) / n * gini(left) + len(right) / n * gini(right)
def best_split(rows, labels):
best = None
for f in range(len(rows[0])):
for t in sorted(set(row[f] for row in rows))[1:]:
left, right = split(rows, labels, f, t)
score = weighted_gini(left, right)
if best is None or score < best[0]:
best = (score, f, t)
return best
def majority(labels):
return max(sorted(set(labels)), key=labels.count)
def build(rows, labels, depth, max_depth, min_samples):
if depth == max_depth or len(rows) < min_samples or gini(labels) == 0:
return {"leaf": majority(labels)}
best = best_split(rows, labels)
if best is None:
return {"leaf": majority(labels)}
score, f, t = best
left_rows, left_labels, right_rows, right_labels = [], [], [], []
for i in range(len(rows)):
if rows[i][f] < t:
left_rows.append(rows[i])
left_labels.append(labels[i])
else:
right_rows.append(rows[i])
right_labels.append(labels[i])
left = build(left_rows, left_labels, depth + 1, max_depth, min_samples)
right = build(right_rows, right_labels, depth + 1, max_depth, min_samples)
if "leaf" in left and "leaf" in right and left["leaf"] == right["leaf"]:
return left # both sides say the same thing: the question is useless
return {"feature": f, "threshold": t, "left": left, "right": right}
def print_rules(node, names, indent=""):
if "leaf" in node:
print(indent + "-> " + node["leaf"])
return
print(indent + "if " + names[node["feature"]] + " < " + str(node["threshold"]) + ":")
print_rules(node["left"], names, indent + " ")
print(indent + "else:")
print_rules(node["right"], names, indent + " ")
def features_used(node, names):
# a leaf uses no feature: return set()
# otherwise: this node's feature name plus everything its two branches use
pass
for depth in [2, 5]:
tree = build(rows, labels, 0, depth, 2)
used = features_used(tree, names)
print(f"depth {depth} uses:", sorted(used))
if "seat_row" in used:
print(" check this: seat_row decides! is that fair?")
راهنما و جواب این چالش · برگشت به درس جلسهی 51: درخت تصمیم ۲
بیشتر بخوان: تابع در پایتون: آموزش def، پارامتر و return با مثال