Explain what machine learning is, how it differs from hand-written rules, and name its three main types with an example of each.

• Definition — a field of AI where systems learn patterns from data, instead of being explicitly programmed with rules.
• Improves with data — the system's performance improves as it sees more data and experience.
• The flow — Data → Learning (algorithms) → Patterns & Predictions → Smart Decisions.
• Three types — Supervised (e.g., spam detection), Unsupervised (e.g., customer segmentation), and Reinforcement Learning (e.g., self-driving cars).
• Reach — powers everything from spam filters to self-driving cars.

Describe what makes learning supervised, its goal, its workflow, and how classification differs from regression.

• Labeled data — the model learns from labeled data — every example already has a known correct answer.
• Goal — learn a mapping from inputs (features) to outputs (labels).
• Classification — predicts categories — discrete outputs like Yes/No, Spam/Not Spam, Cat/Dog. Example: is this email spam or not spam?
• Regression — predicts numbers — continuous values like price, temperature, sales. Example: predicting the price of a house.
• Workflow — labeled data → train the model (learns the relationship between features and labels) → predict / inference on new, unseen data → output.

Describe what unsupervised learning works with, what it's trying to find, its main task, and its benefits.

• No labels — the model finds patterns in data that has no labels or correct answers.
• Goal — discover hidden structure, groupings, or relationships.
• Clustering — the main task — grouping similar items together into clusters or segments.
• Examples — customer segmentation; grouping similar news articles.
• Benefits — no labeled data needed; finds hidden patterns; useful for exploration and segmentation.

Describe how an agent learns, what a policy is, and walk through the reinforcement learning loop.

• Agent — the learner or decision maker; it takes actions in an environment and receives rewards or penalties.
• Goal / policy — learn a strategy (a policy) that maximizes long-term reward through trial and error.
• The loop — agent takes an action → environment responds → reward or penalty (e.g., +10 for a good action, −5 for a bad one) → environment sends the current state back to the agent.
• Examples — game-playing AI, robotics (e.g., pick and place), self-driving car decision-making.
• Benefits — learns optimal behavior through interaction; handles sequential decision-making problems; effective in complex, dynamic environments.

Explain what training is and walk through the five-step training loop, using a model that predicts 4.2 when the correct answer is 7.0.

• Training — showing a model many examples so it can adjust itself to fit the patterns in the data.
• Guess → check → adjust — the model makes a guess, checks how wrong it was (its error/loss), and adjusts.
• Five-step loop — Data (input) → Model prediction → Compare to correct answer → Adjust model (updates its internal parameters to reduce future errors) → Repeat.
• Worked example — prediction 4.2 vs. actual 7.0 → error/loss = 2.8.
• Over time — the loop repeats until predictions get consistently better — from poor predictions / high error to the best model / low error, giving accurate predictions on new, unseen data.

Explain what prediction is, why it matters, and how it works without the correct answer.

• Prediction — once trained, a model takes brand-new, unseen input and produces an output.
• Why it matters — this is the moment a model actually becomes useful in the real world.
• Flow — new, unseen input → trained model analyzes it → predicted label or value.
• No answer needed — it recognizes the patterns it learned during training and gives an instant prediction without needing the correct answer.
• Example — a trained spam model instantly labels a brand-new incoming email (e.g., Not Spam).

Explain spam detection as an ML problem: its type, training data, what it learns, and the five-step flow.

• Problem type — a supervised classification problem: is this email "spam" or "not spam"?
• Training data — thousands of emails already labeled by humans.
• What it learns — patterns in wording, sender behavior, links, and formatting.
• Five steps — training data → the model learns → a new email arrives (unseen) → the model predicts Spam or Not Spam → sorted automatically into the Spam folder or the Inbox.
• Reach — powers the spam folder in nearly every email inbox today.

Explain the data a recommender uses, the mix of techniques, and its five-step flow.

• Inputs — your viewing history, ratings, and similar users' behavior — used to predict what you'll enjoy next.
• Technique mix — a mix of supervised and unsupervised techniques — clustering similar viewers, predicting ratings.
• Five steps — your data → find similar users (clustering) → predict ratings for movies you haven't seen → rank & recommend by predicted rating → you enjoy.
• "Why these?" — you watched similar movies; people with similar taste liked these; high predicted rating for you.
• Reach — powers Netflix, YouTube, and nearly every streaming platform's homepage.

Explain why building a model isn't enough and what evaluation gives you.

• Not the finish line — building a model isn't the finish line — you have to know if it's actually good.
• The trap — a model can look "smart" on training data but fail badly on new, real-world data.
• What evaluation gives — honest, measurable answers about model performance.
• Core metrics — accuracy, precision, recall, F1 score — all calculated from the confusion matrix.

Lay out the confusion matrix and define its four outcomes. Every other metric is built from this table.

• What it is — a table that breaks predictions into 4 outcomes.
• Layout — rows = actual (positive, negative); columns = predicted (positive, negative). Top row: TP | FN. Bottom row: FP | TN.
• TP / TN — true positive = correctly predicted positive; true negative = correctly predicted negative.
• FP / FN — false positive = wrongly predicted positive; false negative = wrongly predicted negative.
• Foundation — every metric — accuracy, precision, recall, F1 — is calculated directly from this table.

Give the accuracy formula, what it measures, and why it can mislead.

• Formula — Accuracy = (TP + TN) / (TP + TN + FP + FN).
• Meaning — the percentage of predictions the model got completely correct.
• Strength — simple and intuitive.
• Weakness — can be misleading on imbalanced data.
• Spam example — a model that always predicts "not spam" can still be 95% accurate if only 5% of emails are truly spam.

Give the precision formula, the question it answers, and when it matters most.

• Formula — Precision = TP / (TP + FP).
• Question — of everything predicted positive, how many were actually correct?
• High precision — very few false alarms.
• When it matters — when false positives are costly — e.g., wrongly flagging a real email as spam.

Give the recall formula, the question it answers, and when it matters most.

• Formula — Recall = TP / (TP + FN).
• Question — of everything that was actually positive, how many did the model catch?
• High recall — very few missed cases.
• When it matters — when false negatives are costly — e.g., missing a genuinely fraudulent transaction.

Give the F1 formula, what it combines, and when it's the right metric.

• Formula — F1 = 2 × (Precision × Recall) / (Precision + Recall).
• What it is — the harmonic mean of precision and recall — balances both into one number.
• When to use — when you care about both false positives and false negatives.
• Imbalanced data — especially valuable on imbalanced datasets, where accuracy alone can mislead.

Explain why cross-validation beats a single train/test split and walk through k-fold with k = 5.

• Purpose — a technique for testing a model's performance more reliably than a single train/test split.
• k-fold — split the data into k parts, train on k−1 parts, test on the remaining part, and repeat k times.
• k = 5 example — 5 rounds; each round a different fold is the test set (Round 1 tests Fold 1 and trains on Folds 2–5, and so on).
• Final estimate — average the performance (the scores) from all rounds.
• Why it matters — reduces the chance of overfitting to a particular split and gives a more stable, trustworthy estimate of performance on new, unseen data.

Describe overfitting, how it performs on training vs. new data, its common causes, and what its curve looks like.

• What it is — the model memorizes the training data too closely, including its noise and randomness.
• Performance — great on training data, but poor on new, unseen data.
• Causes — too complex a model, or too little training data.
• Picture it — an overfit curve wiggles through every training point, bumps and noise included; a model that generalizes well draws a smooth curve that follows the true underlying pattern.
• Takeaway — overfitting means learning the noise instead of the signal; the goal is to generalize well to new, unseen data.

Describe underfitting, how it performs, its common causes, what it looks like, and four ways to fix it.

• What it is — the model is too simple to capture the real pattern in the data.
• Performance — poor on both training data and new data — high error on both.
• Causes — too simple a model, not enough training, or not enough relevant features.
• Picture it — a straight line forced through curved data — it misses important structure.
• What to do — use a more complex model, train longer, add more relevant features, reduce overly strong regularization.