Introduction to Machine Learning: Types, Business Framing and the ML Lifecycle

What is machine learning?

"A computer program is said to learn from experience E with respect to some task T and performance measure P if its performance at T, as measured by P, improves with experience E." (Tom Mitchell)

Instead of writing rules by hand, we give the computer examples (data) and let an algorithm learn the rules (a model) that map inputs to outputs.

They overlap heavily: linear and logistic regression belong to both.

Types of machine learning

import pandas as pd
h = pd.read_csv("house_prices.csv"); tc = pd.read_csv("telecom_churn.csv")
print("Regression target:", "price_lakh", "| examples:", h.price_lakh.head(3).tolist(), "| features:", [c for c in h.columns if c not in ("house_id", "price_lakh")])
print("Classification target: churn", tc.churn.value_counts().to_dict(), "| features:", [c for c in tc.columns if c not in ("customer_id", "churn")])

Regression target: price_lakh | examples: [60.2, 60.9, 43.7] | features: ['city', 'area_sqft', 'bedrooms', 'bathrooms', 'age_years', 'distance_to_center_km', 'parking']
Classification target: churn {'No': 2228, 'Yes': 772} | features: ['senior_citizen', 'tenure_months', 'contract', 'internet_service', 'payment_method', 'monthly_charges', 'total_charges', 'support_calls']

Vocabulary: a row is an observation/sample; input columns are features (X, predictors, independent variables); the column to predict is the target (y, label, dependent variable); the learned function is the model; using it on new data is inference/prediction.

From business problem to ML problem

When not to use ML

Simple rules work; there is little data or no labels; mistakes are unacceptable and must be explained exactly; the pattern changes faster than you can retrain; or nobody will act on the predictions. Always compare with a simple baseline.

The ML lifecycle

In practice, most time goes to framing, data preparation and deployment, not to choosing algorithms.

Linear vs non-linear models and link functions

A linear model predicts with a weighted sum of features:

Generalised linear models pass that linear score through a link function so the output fits the target type:

Non-linear models (decision trees, random forests, gradient boosting, KNN, neural networks) capture thresholds and interactions automatically, usually at the cost of interpretability.

Overfitting and generalisation (preview)

A model must work on new data. Memorising the training set (overfitting) fails in production; being too simple (underfitting) misses real patterns. We therefore always evaluate on data the model has not seen (lessons 8 and 15).

Where does GenAI fit?

Large language models are trained with self-supervised learning on huge text corpora, then adapted with instruction tuning and reinforcement learning from human feedback. In analytics they complement classical ML: LLMs read and write text, code and SQL; classical models remain best for structured, tabular prediction (churn, credit risk, demand), where gradient boosting usually wins. The GenAI track covers LLMs, RAG and agents.

Summary