Most modern applications you use today, from recommendations on Netflix to spam filters in your email, relies on systems that improve with data. Machine Learning is the field that makes this possible.

Instead of explicitly programming every rule, we train models to learn patterns from data and make decisions or predictions. Whether it’s recognizing images, understanding text, or forecasting trends, machine learning shifts the focus from writing rules to learning from examples.

This chapter covers what machine learning actually is, how it differs from traditional programming, the main types of ML, and where it makes sense to use it (and where it doesn't).

Traditional programming and machine learning solve problems in fundamentally different ways.

In traditional programming, a developer studies the problem, figures out the rules, and writes code that implements those rules. The computer is just following instructions. Input goes in, rules are applied, output comes out.

Machine learning flips this. You give the computer inputs along with the correct outputs (examples), and it discovers the rules by itself. Those learned rules are what we call a "model." Once trained, the model can take new inputs it has never seen before and produce reasonable outputs.

The key difference: in traditional programming, the developer encodes domain knowledge as logic. In ML, the developer encodes domain knowledge as training data, and the algorithm extracts the logic.

So what does "learning" actually mean here?

At its core, the model has internal parameters (think of them as millions of adjustable knobs). During training, the algorithm tweaks these knobs to minimize the gap between what the model predicts and the correct answers in the training data.

Over thousands of iterations, the model gets better at capturing the patterns that matter. Once the knobs are set, the model can generalize to new data it has never seen.

This is both the power and the limitation of ML. The model can only learn patterns that exist in the training data. If your house pricing data only includes sales from California, the model won't know how to price homes in Texas.

With the core idea of learning from data in place, the natural question is: what kind of data, and what kind of learning?

There are three main paradigms, each suited to different problems.

Supervised learning

Supervised learning is the most common type in production systems. You give the model labeled examples, input-output pairs where the correct answer is known. The model learns to map inputs to outputs.

Example: Email labeled spam or not-spam, houses labeled with sale prices, images labeled with object names. The "supervised" part means that you need to provide the right answers during training.

Unsupervised learning

Unsupervised learning works with unlabeled data. There are no labels or correct answers. Instead, the model finds hidden structure on its own. Clustering is the classic example: given a million customers with purchase histories, group them into segments that share similar behaviors. Nobody told the model what the segments should be.

Other examples include anomaly detection (finding data points that don't fit any pattern) and dimensionality reduction (compressing high-dimensional data into something manageable). Unsupervised learning is valuable when you don't know what patterns exist yet and want the data to tell you.

Reinforcement learning

Reinforcement learning takes a completely different approach. An agent interacts with an environment, takes actions, and receives rewards or penalties. Over time, it learns a strategy that maximizes cumulative reward.

Think of it like training a dog: good behavior gets a treat, bad behavior doesn't. This is how AlphaGo learned to play Go, how robots learn to walk, and how self-driving cars learn driving policies.

Reinforcement learning shows up less often in typical ML system design interviews, but it is important to know it exists, especially for problems involving sequential decision-making.

Here is a quick comparison of the three types:

We will cover supervised and unsupervised learning in much more depth in the next section of this course. For now, the key takeaway is that the type of learning you choose depends on what data you have and what problem you are solving.

Not every problem needs machine learning. This might seem obvious, but it is one of the most common mistakes in practice and in interviews. Reaching for ML when a simpler approach works is a red flag.

The core principle: ML is the right tool when the rules governing a problem are too complex, too numerous, or change too fast for humans to write and maintain. Traditional software is better when the logic is well-defined, deterministic, and stable.

Some problems are better solved with simple logic. Calculating sales tax is a good example. Rates are defined and rarely ambiguous, so a lookup table works perfectly. Adding ML would only make the system harder to maintain and less reliable.

Other problems don’t fit clean rules. Fraud detection is one of them. Patterns change constantly, and static rules miss new attacks. ML models can pick up subtle signals like unusual transaction patterns, location mismatches, or device behavior, and adapt as those patterns evolve.

In practice, production systems combine both. An ML model might assign a risk score, while hard rules enforce strict constraints, like blocking transactions from sanctioned regions. The model handles ambiguity. The rules provide guardrails.

There’s also a cost to ML. It requires data pipelines, training infrastructure, serving systems, and ongoing monitoring. If a problem can be solved with simple logic, that’s usually the better choice. Use ML when the complexity truly demands it.

To make this concrete, here are four domains where ML has become essential, and in each case, why rules alone would not work.

Recommendation Engines

When you open Netflix, YouTube, or Spotify, the content you see is personalized. With millions of items in the catalog and millions of users, each with unique preferences, no team of curators could manually match users to content. Showing everyone the most popular items is a decent baseline, but it treats every user identically. A jazz fan and a hip-hop fan opening Spotify should see entirely different playlists.

ML models learn individual preferences from behavioral signals: what you watched, how long you watched it, what you skipped, what you searched for. They combine these signals with item features (genre, creator, recency) to predict what you are most likely to engage with next.

The following diagram shows a simplified version of how recommendations work:

The system first generates a broad set of candidates (hundreds from millions) using a fast, lightweight model. Then a more powerful ranking model scores those candidates and orders them by predicted relevance. This two-stage approach is a recurring pattern across ML systems.

Search Ranking

A query like "running shoes" on Amazon returns thousands of results. Which ones appear on the first page determines whether a user finds what they want or leaves frustrated. Simple keyword matching gets you a baseline, but it does not account for relevance, user intent, product quality, or personalization.

ML ranking models take in dozens of features: how well the query matches the product title and description, the product's rating and review count, the user's purchase history and browsing behavior, the product's sales velocity, and more. The model learns which combinations of these features predict a click or purchase. Google's search ranking uses hundreds of such signals, continuously updated as user behavior evolves.

Fraud Detection

Banks and payment processors handle millions of transactions per second. Fraudsters are creative and constantly change tactics. A rule that catches today's fraud pattern might be useless next month.

ML models learn subtle signals that rules cannot easily capture. A transaction from a new device in an unusual location, at an unusual time, for an unusual amount, all within seconds of a password change, is suspicious. Any single signal might be innocent. The combination is what matters, and ML models excel at learning these multi-factor patterns. They also adapt as new fraud techniques emerge, because the model can be retrained on recent labeled examples.

Self-Driving Cars

Driving involves thousands of simultaneous decisions in a constantly changing environment. No set of if-else rules can anticipate every combination of road conditions, pedestrian behavior, weather, and traffic patterns.

Self-driving systems use multiple ML models working together. Perception models detect objects from camera and lidar data: other cars, pedestrians, lane markings, traffic signs. Prediction models forecast what other road users will do in the next few seconds. Planning models decide what action the car should take given the current state and predictions. This is an example of multiple ML models composing a larger system, each solving a piece of the overall problem.

Here is a summary of these four applications:

In each case, the pattern is the same: the problem involves too many variables, too many combinations, or too much change for hand-written rules to keep up. That is where ML shines.