🤖 Week 7 – Machine Learning Core · Day 36
Day 36: What is Machine Learning?
⏱ 90 min📶 Beginner🎯 100% self-study
Day 36 / 60
✅ Today you will learn
- Rules vs learning from examples
- Supervised, unsupervised, reinforcement
- Regression vs classification
- Features, labels, training and testing
- Overfitting and the 7-step ML workflow
🏢 In the real world
A bank used to approve loans with a rule book: “income above ₹40,000 and credit score above 700”. It missed good customers and approved risky ones. With 1,200 past applications and their outcomes, a machine-learning model learns the pattern itself – and keeps improving as new data arrives.
Rules vs learning
Traditional programming
- You write the rules
- Data + rules → answers
- Breaks when the world changes
- Example: if score > 700 then approve
Machine learning
- The computer finds the rules
- Data + answers → rules (model)
- Retrain when data changes
- Example: model learns from 1,200 past loans
Python
import pandas as pd
loans = pd.read_csv("loan_applications.csv")
print(loans.head())
rule = (loans["monthly_income"] > 40000) & (loans["credit_score"] > 700)
print(f"Old rule accuracy: {(rule.astype(int) == loans['approved']).mean():.1%}")
print(f"Approval rate in history: {loans['approved'].mean():.1%}") applicant_id monthly_income credit_score loan_amount employment years_employed approved
0 L5001 45300 819 502000 Salaried 12 1
1 L5002 61000 799 806000 Self-Employed 25 1
2 L5003 71200 801 367000 Self-Employed 3 1
3 L5004 35000 775 611000 Salaried 19 1
4 L5005 76500 737 2083000 Salaried 27 0
Old rule accuracy: 72.2%
Approval rate in history: 46.2%
Python
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier
loans = pd.read_csv("loan_applications.csv")
X = loans[["monthly_income", "credit_score", "loan_amount", "years_employed"]]
y = loans["approved"]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)
model = DecisionTreeClassifier(max_depth=4, random_state=42).fit(X_train, y_train)
print(f"ML model accuracy on unseen applications: {model.score(X_test, y_test):.1%}")ML model accuracy on unseen applications: 81.0%
🎉 Your first model
Five lines of scikit-learn beat the hand-written rule on applications the model has never seen. Tomorrow you will understand each line properly.
Types of machine learning
🏷️
SupervisedLearn from examples with answers (labels). Loan approved? House price?🧩
UnsupervisedFind structure without labels – customer segments, anomalies.🎮
ReinforcementLearn by trial and reward – games, robots, ad bidding.| Problem | Type | Output | Course day |
|---|---|---|---|
| Predict house price | Supervised – regression | a number (₹ lakh) | 38–39 |
| Approve a loan? | Supervised – classification | a class (yes/no) | 40–43 |
| Will a customer churn? | Supervised – classification | probability | 56–60 |
| Group mall customers | Unsupervised – clustering | segment id | 46 |
| Compress many columns | Unsupervised – dimensionality reduction | fewer features | 47 |
| Review positive/negative | Supervised – NLP classification | sentiment | 48 |
| Next month’s AQI | Time-series forecasting | future values | 49 |
Key words
Python
import pandas as pd
h = pd.read_csv("house_prices.csv")
X = h.drop(columns="price_lakh") # features (inputs)
y = h["price_lakh"] # label / target (what we predict)
print("Features:", list(X.columns))
print("Target:", y.name, "| one example →", dict(X.iloc[0]), "→", y.iloc[0], "lakh")Features: ['city', 'area_sqft', 'bedrooms', 'bathrooms', 'age_years', 'metro_km', 'furnishing', 'parking']
Target: price_lakh | one example → {'city': 'Pune', 'area_sqft': np.int64(1210), 'bedrooms': np.int64(2), 'bathrooms': np.int64(2), 'age_years': np.int64(18), 'metro_km': np.float64(7.0), 'furnishing': 'Furnished', 'parking': 'No'} → 55.07 lakh
| Word | Meaning |
|---|---|
| Feature (X) | An input column: area, bedrooms, city |
| Label / target (y) | The answer we want to predict: price |
| Model | The learned function f(X) ≈ y |
| Training | Fitting the model on known examples |
| Test set | Held-back examples to check the model honestly |
| Prediction | Model output for new X |
| Overfitting | Memorising training data, failing on new data |
Overfitting – memorising vs learning
Python
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier
loans = pd.read_csv("loan_applications.csv")
X = loans[["monthly_income", "credit_score", "loan_amount", "years_employed"]]; y = loans["approved"]
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)
for depth in [1, 2, 4, 8, 16, None]:
m = DecisionTreeClassifier(max_depth=depth, random_state=42).fit(X_tr, y_tr)
print(f"depth={str(depth):4s} train={m.score(X_tr, y_tr):.3f} test={m.score(X_te, y_te):.3f}")depth=1 train=0.787 test=0.817
depth=2 train=0.787 test=0.817
depth=4 train=0.829 test=0.810
depth=8 train=0.922 test=0.783
depth=16 train=1.000 test=0.743
depth=None train=1.000 test=0.743
Python
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier
loans = pd.read_csv("loan_applications.csv")
X = loans[["monthly_income", "credit_score", "loan_amount", "years_employed"]]; y = loans["approved"]
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)
d = range(1, 16); tr, te = [], []
for k in d:
m = DecisionTreeClassifier(max_depth=k, random_state=42).fit(X_tr, y_tr)
tr.append(m.score(X_tr, y_tr)); te.append(m.score(X_te, y_te))
plt.figure(figsize=(8.5, 3.8))
plt.plot(d, tr, marker="o", label="training accuracy", color="#94a3b8")
plt.plot(d, te, marker="o", label="test accuracy", color="#4f46e5")
best = d[te.index(max(te))]; plt.axvline(best, color="#16a34a", ls="--", label=f"sweet spot (depth {best})")
plt.xlabel("model complexity (tree depth)"); plt.ylabel("accuracy"); plt.legend()
plt.title("Too simple = underfit, too complex = overfit"); plt.show()
📚 Exam analogy
A student who memorises last year’s paper scores 100% on it but fails a new paper. The test set is the “new paper” – never let the model see it during training.
The ML workflow
- 1Define the business problem and metric
- 2Collect and clean data
- 3EDA and feature engineering
- 4Split train / test
- 5Train baseline, then better models
- 6Evaluate on test set
- 7Deploy, monitor, retrain
⚠️ ML is not always the answer
If a simple rule or SQL report solves the problem, use it. ML needs enough good historical data with labels, a clear metric, and someone to maintain it.
🏅 Job skills unlocked today
✔ Explain ML vs traditional programming✔ Classify problems as regression, classification or clustering✔ Use the words features, target, train, test✔ Recognise overfitting
Practice tasks
🏋️ Practice – do it yourself
Task 1: Classify each problem: (a) predict tomorrow’s Delhi AQI, (b) detect spam SMS, (c) group shoppers by behaviour, (d) estimate a used car’s price.
Task 2: Train the decision tree with only
credit_score and monthly_income. How much accuracy is lost?Show solution
Python
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier
l = pd.read_csv("loan_applications.csv")
X_tr, X_te, y_tr, y_te = train_test_split(l[["credit_score", "monthly_income"]], l["approved"], test_size=0.25, random_state=42, stratify=l["approved"])
print(round(DecisionTreeClassifier(max_depth=4, random_state=42).fit(X_tr, y_tr).score(X_te, y_te), 3))0.783
Task 3: List the features and the target for a model that predicts student average marks.
Quick quiz
🧠 Quick Quiz – test yourself
1. Predicting a house price is…
The output is a number.
2. Grouping customers without labels is…
Clustering is unsupervised.
3. Training accuracy 100%, test 70% means…
It memorised the training data.
4. The test set is used to…
Keep it untouched until the end.
5. In scikit-learn, y is…
X = features, y = target.
Interview corner
💼 Interview Corner
What is the difference between supervised and unsupervised learning?
Supervised learning trains on labelled examples to predict a known target (price, churn). Unsupervised learning finds structure in unlabelled data (clusters, components, anomalies).
What is overfitting and how do you prevent it?
The model learns noise in the training data and fails on new data. Prevent it with a held-out test set, cross-validation, simpler models or regularisation, more data, and early stopping.
Summary
- ML learns rules from examples instead of hand-coding them.
- Supervised (regression, classification), unsupervised, reinforcement.
- X = features, y = target; always evaluate on unseen test data.
- Overfitting = memorising; aim for the best test score.
➡️ Next
Day 37: scikit-learn step by step – fit, predict, score, and a proper baseline.
Frequently Asked Questions
Do I need advanced maths for ML?
For using ML well you need intuition about statistics (Days 26–31) more than calculus. The maths helps later for research roles.
Is AI the same as ML?
AI is the broad goal of intelligent machines. ML is the main way we build AI today; deep learning (Day 51) is a part of ML.
