Confusing Data Science Terms Explained in Simple Words

import statistics

monthly_spend = [800, 950, 1000, 1100, 1200, 40000]   # one very large customer
print("mean:  ", round(statistics.mean(monthly_spend)))
print("median:", statistics.median(monthly_spend))
# The mean is pulled up by one customer; the median shows a typical one.

import numpy as np
from scipy import stats

rng = np.random.default_rng(0)
a = rng.normal(100.0, 15, 200_000)
b = rng.normal(100.3, 15, 200_000)   # tiny true difference

t, p = stats.ttest_ind(a, b)
print(f"p-value: {p:.2e}")                       # very small -> 'significant'
print(f"difference in means: {b.mean() - a.mean():.2f}")  # but tiny effect
# Always report the effect size, not just the p-value.

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier

X, y = make_classification(n_samples=800, n_features=8, flip_y=0.1, random_state=1)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, random_state=1)

for depth, label in [(1, "underfit"), (4, "balanced"), (None, "overfit")]:
    m = DecisionTreeClassifier(max_depth=depth, random_state=1).fit(X_tr, y_tr)
    print(f"{label:9} train={m.score(X_tr, y_tr):.2f} test={m.score(X_te, y_te):.2f}")

from sklearn.metrics import accuracy_score, recall_score, balanced_accuracy_score

# 100 transactions, 2 are fraud; the model always predicts 'not fraud'
y_true = [1, 1] + [0] * 98
y_pred = [0] * 100

print("accuracy:         ", accuracy_score(y_true, y_pred))          # 0.98
print("recall (fraud):   ", recall_score(y_true, y_pred))            # 0.0
print("balanced accuracy:", balanced_accuracy_score(y_true, y_pred)) # 0.5

import numpy as np
from sklearn.preprocessing import MinMaxScaler, StandardScaler

spend = np.array([[200], [500], [800], [1100], [5000]])
print("normalised (0-1):", MinMaxScaler().fit_transform(spend).ravel().round(2))
print("standardised:    ", StandardScaler().fit_transform(spend).ravel().round(2))
# Fit scalers on training data only, then apply to validation and test.

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier

X, y = make_classification(n_samples=1000, random_state=3)
X_tmp, X_test, y_tmp, y_test = train_test_split(X, y, test_size=0.2, random_state=3)
X_tr, X_val, y_tr, y_val = train_test_split(X_tmp, y_tmp, test_size=0.25, random_state=3)

best = max([2, 4, 6, 8], key=lambda d: DecisionTreeClassifier(max_depth=d, random_state=3)
           .fit(X_tr, y_tr).score(X_val, y_val))       # tune on validation
final = DecisionTreeClassifier(max_depth=best, random_state=3).fit(X_tmp, y_tmp)
print("chosen depth:", best, "test score:", round(final.score(X_test, y_test), 3))

Correlation vs causation  -> 'Together' is not 'because'
p-value vs effect size    -> Surprising is not the same as big
Bias vs variance          -> Too simple vs too jumpy
Precision vs recall       -> Purity of flags vs coverage of cases
Validation vs test        -> Practice test vs final exam
Normalise vs standardise  -> Squeeze into a range vs centre and scale

Pair: ____________ vs ____________
One-line difference: ______________________
Everyday example: ______________________
Code check: ______________________
When the mix-up causes harm: ______________________

Answer in one sentence each:

1. Why does correlation not prove causation?
2. Can a result be significant but unimportant?
3. What does a big train-test gap tell you?
4. When does recall matter more than precision?
5. Why keep the test set untouched until the end?
6. When might one-hot encoding be better than label encoding?