Applications of Data Analytics — Bringing It All Together

This final lesson revisits applications with everything you have learned across the four units, tracing complete end-to-end pipelines and showing exactly which technique from which unit does which job.

Case Study 1 — E-Commerce Recommendation and Retention

Business problem: increase repeat purchases and average order value.

# A compact version of the segmentation + rules pipeline
import pandas as pd, numpy as np
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans

np.random.seed(42)
customers = pd.DataFrame({
    "customer_id": [f"C{i:04d}" for i in range(300)],
    "recency_days": np.random.randint(1, 200, 300),
    "frequency": np.random.poisson(4, 300) + 1,
    "monetary": np.random.gamma(3, 4000, 300).round(0),
})

X = StandardScaler().fit_transform(customers[["recency_days","frequency","monetary"]])
customers["segment"] = KMeans(n_clusters=4, n_init=10, random_state=42).fit_predict(X)

profile = customers.groupby("segment")[["recency_days","frequency","monetary"]].mean().round(1)
profile["size"] = customers["segment"].value_counts().sort_index()
profile["strategy"] = [
    "Re-engagement campaign", "VIP loyalty programme",
    "Upsell higher-value items", "Win-back discount",
]
print(profile)

Case Study 2 — Healthcare: Early Disease Risk Prediction

The single most important design decision here is the metric. As shown in the classification lesson, a 97%-accurate model that misses 40% of real cases is unacceptable in medicine. Optimise for recall, and use Bayes' theorem (Unit 2) to explain to clinicians why a positive screening result still needs a confirmatory test.

Case Study 3 — Banking: Real-Time Fraud Detection

Case Study 4 — Education: Predicting At-Risk Students

import pandas as pd, numpy as np
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.metrics import classification_report, confusion_matrix

np.random.seed(42)
n = 600
students = pd.DataFrame({
    "attendance_pct":  np.random.beta(6, 2, n) * 100,
    "assignments_done": np.random.binomial(10, 0.75, n),
    "internal_marks":  np.random.normal(60, 15, n).clip(0, 100),
    "lms_logins_week": np.random.poisson(6, n),
})
# True underlying risk
risk = (-0.06*students["attendance_pct"] - 0.30*students["assignments_done"]
        - 0.05*students["internal_marks"] - 0.12*students["lms_logins_week"] + 11)
students["will_fail"] = (risk + np.random.normal(0, 0.8, n) > 0).astype(int)

X = students.drop(columns="will_fail")
y = students["will_fail"]
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25,
                                          random_state=42, stratify=y)

model = make_pipeline(StandardScaler(),
                      LogisticRegression(class_weight="balanced", max_iter=2000))
model.fit(X_tr, y_tr)
pred = model.predict(X_te)

print(confusion_matrix(y_te, pred))
print(classification_report(y_te, pred, target_names=["Pass", "At Risk"]))

coefs = pd.DataFrame({
    "feature": X.columns,
    "coefficient": model.named_steps["logisticregression"].coef_[0].round(3),
}).sort_values("coefficient")
print("\nRisk drivers (most protective first):\n", coefs)
# Intervention: flag high-risk students by week 8 for mentoring —
# recall matters more than precision, since a false alarm just means
# an extra counselling session.

Domain Summary Table

The Analyst's Toolkit — Complete Map

Ethics and Responsible Analytics

Career Paths

Course Wrap-Up — How It All Connects

Every concept in this course is one link in that loop. The statistics of Unit 2 are what let you tell a real pattern from noise in Unit 3. The cleaning discipline of Unit 1 is what makes any of Unit 3's models trustworthy. And the tools of Unit 4 are what let you do all of it on data that a spreadsheet could never open.

The most valuable habit you can carry out of this course: always plot the data before you model it, always report the effect size alongside the p-value, and never let a model's confidence exceed the quality of the data it was trained on.